├── LICENSE ├── README.md ├── adv_examples └── L2 │ └── f3 │ └── all │ └── 0.3 confidence │ ├── Best example of 0 Distortion 103.01087188720703.png │ ├── Best example of 1 Distortion 105.97237396240234.png │ ├── Best example of 2 Distortion 94.40467834472656.png │ ├── Best example of 3 Distortion 85.84400939941406.png │ ├── Best example of 4 Distortion 101.89212036132812.png │ ├── Best example of 5 Distortion 99.68110656738281.png │ ├── Best example of 6 Distortion 75.24471282958984.png │ ├── Best example of 7 Distortion 113.2362289428711.png │ ├── Best example of 8 Distortion 84.97225189208984.png │ ├── Best example of 9 Distortion 100.87386322021484.png │ ├── Daedalus example batch.npz │ └── Distortions of images 0 to 10.npy ├── data └── coco_classes.txt ├── demo_plot.py ├── l0_yolov3.py ├── l2_ensemble.py ├── l2_retinanet.py ├── l2_yolov3.py ├── model ├── darknet53.py └── yolo_model.py └── resources └── l2attack.jpg /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Daedalus-attack 2 | The code of our paper "Daedalus: Breaking Non-Maximum Suppression in Object Detection via Adversarial Examples". 3 | 4 | We propose an attack, in which we can tune the strength of the attack and specify the object category to attack, to break non-maximum suppression (NMS) in object detection. As the consequence, the detection model outputs extremely dense results as redundant detection boxes are not filtered by NMS. 5 | 6 | Some results are displayed here: 7 | ![Alt text](resources/l2attack.jpg) 8 | *Adversarial examples made by our L2 attack. The first row contains original images. The third row contains our low-confidence (0.3) adversarial examples. The fifth row contains our high-confidence (0.7) examples. The detection results from YOLO-v3 are in the rows below them. The confidence controls the density of the redundant detection boxes in the detection results.* 9 | 10 | **Launching real-world attacks via a Daedalus poster** 11 | 12 | We instantiated the Daedalus perturbation into a physical poster. You can watch the demo of the attack on YouTube: 13 | [![Watch the video](https://img.youtube.com/vi/U1LsTl8vufM/maxresdefault.jpg)](https://www.youtube.com/watch?v=U1LsTl8vufM) 14 | The code for generating posters against YOLO-v3 is in [this](https://github.com/NeuralSec/Daedalus-physical) repository (for academic purpose only). 15 | 16 | --- 17 | 18 | **Running the attack against YOLO-v3:** 19 | 20 | 1. Download [yolo.h5](https://1drv.ms/u/s!AqftEu9YAdEGidZ7vEm-4v4c2sV-Lw) and put it into '../model'; 21 | 2. Put original images into '../Datasets/COCO/val2017/'; 22 | 3. Run l2_yolov3.py. 23 | 24 | **Running the attack against RetinaNet:** 25 | 26 | 1. Install [keras-retinanet](https://github.com/fizyr/keras-retinanet); 27 | 2. Download [resnet50_coco_best_v2.1.0.h5](https://drive.google.com/file/d/1N6Xg5SOW8Ic4hpC8PoIRvggcstx0HcXw/view?usp=sharing) and put it into '../model'; 28 | 3. Put original images into '../Datasets/COCO/val2017/'; 29 | 4. Run l2_retinanet.py. 30 | 31 | **Running ensemble attack to craft robust adversarial examples:** 32 | 33 | Run l2_ensemble.py after completing the above setups for YOLO-v3 and RetinaNet attacks. 34 | 35 | All attacks can specify object categories to attack. Crafted adversarial examples will be stored as 416X416 sized .png files in '../adv_examples/...'. The examples can be tested on official darknet and retinanet. 36 | 37 | Cite this work: 38 | 39 | ``` 40 | @article{wang2021daedalus, 41 | title={Daedalus: Breaking nonmaximum suppression in object detection via adversarial examples}, 42 | author={Wang, Derui and Li, Chaoran and Wen, Sheng and Han, Qing-Long and Nepal, Surya and Zhang, Xiangyu and Xiang, Yang}, 43 | journal={IEEE Transactions on Cybernetics}, 44 | volume={52}, 45 | number={8}, 46 | pages={7427--7440}, 47 | year={2021}, 48 | publisher={IEEE} 49 | } 50 | ``` 51 | -------------------------------------------------------------------------------- /adv_examples/L2/f3/all/0.3 confidence/Best example of 0 Distortion 103.01087188720703.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/NeuralSec/Daedalus-attack/3f9cb38f6389cb7cd4895a2f9679aa5ce3c81e70/adv_examples/L2/f3/all/0.3 confidence/Best example of 0 Distortion 103.01087188720703.png -------------------------------------------------------------------------------- /adv_examples/L2/f3/all/0.3 confidence/Best example of 1 Distortion 105.97237396240234.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/NeuralSec/Daedalus-attack/3f9cb38f6389cb7cd4895a2f9679aa5ce3c81e70/adv_examples/L2/f3/all/0.3 confidence/Best example of 1 Distortion 105.97237396240234.png -------------------------------------------------------------------------------- /adv_examples/L2/f3/all/0.3 confidence/Best example of 2 Distortion 94.40467834472656.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/NeuralSec/Daedalus-attack/3f9cb38f6389cb7cd4895a2f9679aa5ce3c81e70/adv_examples/L2/f3/all/0.3 confidence/Best example of 2 Distortion 94.40467834472656.png -------------------------------------------------------------------------------- /adv_examples/L2/f3/all/0.3 confidence/Best example of 3 Distortion 85.84400939941406.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/NeuralSec/Daedalus-attack/3f9cb38f6389cb7cd4895a2f9679aa5ce3c81e70/adv_examples/L2/f3/all/0.3 confidence/Best example of 3 Distortion 85.84400939941406.png -------------------------------------------------------------------------------- /adv_examples/L2/f3/all/0.3 confidence/Best example of 4 Distortion 101.89212036132812.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/NeuralSec/Daedalus-attack/3f9cb38f6389cb7cd4895a2f9679aa5ce3c81e70/adv_examples/L2/f3/all/0.3 confidence/Best example of 4 Distortion 101.89212036132812.png -------------------------------------------------------------------------------- /adv_examples/L2/f3/all/0.3 confidence/Best example of 5 Distortion 99.68110656738281.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/NeuralSec/Daedalus-attack/3f9cb38f6389cb7cd4895a2f9679aa5ce3c81e70/adv_examples/L2/f3/all/0.3 confidence/Best example of 5 Distortion 99.68110656738281.png -------------------------------------------------------------------------------- /adv_examples/L2/f3/all/0.3 confidence/Best example of 6 Distortion 75.24471282958984.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/NeuralSec/Daedalus-attack/3f9cb38f6389cb7cd4895a2f9679aa5ce3c81e70/adv_examples/L2/f3/all/0.3 confidence/Best example of 6 Distortion 75.24471282958984.png -------------------------------------------------------------------------------- /adv_examples/L2/f3/all/0.3 confidence/Best example of 7 Distortion 113.2362289428711.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/NeuralSec/Daedalus-attack/3f9cb38f6389cb7cd4895a2f9679aa5ce3c81e70/adv_examples/L2/f3/all/0.3 confidence/Best example of 7 Distortion 113.2362289428711.png -------------------------------------------------------------------------------- /adv_examples/L2/f3/all/0.3 confidence/Best example of 8 Distortion 84.97225189208984.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/NeuralSec/Daedalus-attack/3f9cb38f6389cb7cd4895a2f9679aa5ce3c81e70/adv_examples/L2/f3/all/0.3 confidence/Best example of 8 Distortion 84.97225189208984.png -------------------------------------------------------------------------------- /adv_examples/L2/f3/all/0.3 confidence/Best example of 9 Distortion 100.87386322021484.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/NeuralSec/Daedalus-attack/3f9cb38f6389cb7cd4895a2f9679aa5ce3c81e70/adv_examples/L2/f3/all/0.3 confidence/Best example of 9 Distortion 100.87386322021484.png -------------------------------------------------------------------------------- /adv_examples/L2/f3/all/0.3 confidence/Daedalus example batch.npz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/NeuralSec/Daedalus-attack/3f9cb38f6389cb7cd4895a2f9679aa5ce3c81e70/adv_examples/L2/f3/all/0.3 confidence/Daedalus example batch.npz -------------------------------------------------------------------------------- /adv_examples/L2/f3/all/0.3 confidence/Distortions of images 0 to 10.npy: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/NeuralSec/Daedalus-attack/3f9cb38f6389cb7cd4895a2f9679aa5ce3c81e70/adv_examples/L2/f3/all/0.3 confidence/Distortions of images 0 to 10.npy -------------------------------------------------------------------------------- /data/coco_classes.txt: -------------------------------------------------------------------------------- 1 | person 2 | bicycle 3 | car 4 | motorbike 5 | aeroplane 6 | bus 7 | train 8 | truck 9 | boat 10 | traffic light 11 | fire hydrant 12 | stop sign 13 | parking meter 14 | bench 15 | bird 16 | cat 17 | dog 18 | horse 19 | sheep 20 | cow 21 | elephant 22 | bear 23 | zebra 24 | giraffe 25 | backpack 26 | umbrella 27 | handbag 28 | tie 29 | suitcase 30 | frisbee 31 | skis 32 | snowboard 33 | sports ball 34 | kite 35 | baseball bat 36 | baseball glove 37 | skateboard 38 | surfboard 39 | tennis racket 40 | bottle 41 | wine glass 42 | cup 43 | fork 44 | knife 45 | spoon 46 | bowl 47 | banana 48 | apple 49 | sandwich 50 | orange 51 | broccoli 52 | carrot 53 | hot dog 54 | pizza 55 | donut 56 | cake 57 | chair 58 | sofa 59 | pottedplant 60 | bed 61 | diningtable 62 | toilet 63 | tvmonitor 64 | laptop 65 | mouse 66 | remote 67 | keyboard 68 | cell phone 69 | microwave 70 | oven 71 | toaster 72 | sink 73 | refrigerator 74 | book 75 | clock 76 | vase 77 | scissors 78 | teddy bear 79 | hair drier 80 | toothbrush 81 | -------------------------------------------------------------------------------- /demo_plot.py: -------------------------------------------------------------------------------- 1 | import os 2 | import cv2 3 | import numpy as np 4 | import matplotlib.pyplot as plt 5 | import matplotlib.gridspec as gridspec 6 | 7 | PATHS = ['adv_examples/plot/L2/f3/benign', 8 | 'adv_examples/plot/L2/f3/0.3', 9 | 'adv_examples/plot/L2/f3/0.7'] 10 | 11 | RESULT_PATH = ['adv_examples/plot/L2/f3/benignresult', 12 | 'adv_examples/plot/L2/f3/0.3result', 13 | 'adv_examples/plot/L2/f3/0.7result'] 14 | 15 | 16 | def plot(paths, result_paths): 17 | b_imgs = [] 18 | b_r = [] 19 | low_imgs = [] 20 | l_r = [] 21 | high_imgs =[] 22 | h_r = [] 23 | for (root, dirs, files) in os.walk(paths[0]): 24 | if files: 25 | for f in files: 26 | path = os.path.join(root, f) 27 | originalimgs = cv2.imread(path) # RGB image 28 | originalimgs = cv2.cvtColor(originalimgs, cv2.COLOR_BGR2RGB) 29 | originalimgs = cv2.resize(originalimgs, (416, 416), interpolation=cv2.INTER_CUBIC) 30 | b_imgs.append(originalimgs) 31 | for (root, dirs, files) in os.walk(result_paths[0]): 32 | if files: 33 | for f in files: 34 | path = os.path.join(root, f) 35 | originalimgs = cv2.imread(path) # RGB image 36 | originalimgs = cv2.cvtColor(originalimgs, cv2.COLOR_BGR2RGB) 37 | originalimgs = cv2.resize(originalimgs, (416, 416), interpolation=cv2.INTER_CUBIC) 38 | b_r.append(originalimgs) 39 | 40 | for (root, dirs, files) in os.walk(paths[1]): 41 | if files: 42 | for f in files: 43 | path = os.path.join(root, f) 44 | lowconfimgs = cv2.imread(path) # RGB image 45 | lowconfimgs = cv2.cvtColor(lowconfimgs, cv2.COLOR_BGR2RGB) 46 | low_imgs.append(lowconfimgs) 47 | for (root, dirs, files) in os.walk(result_paths[1]): 48 | if files: 49 | for f in files: 50 | path = os.path.join(root, f) 51 | lowconfimgs = cv2.imread(path) # RGB image 52 | lowconfimgs = cv2.cvtColor(lowconfimgs, cv2.COLOR_BGR2RGB) 53 | l_r.append(lowconfimgs) 54 | 55 | for (root, dirs, files) in os.walk(paths[2]): 56 | if files: 57 | for f in files: 58 | path = os.path.join(root, f) 59 | highconfimgs = cv2.imread(path) # RGB image 60 | highconfimgs = cv2.cvtColor(highconfimgs, cv2.COLOR_BGR2RGB) 61 | high_imgs.append(highconfimgs) 62 | for (root, dirs, files) in os.walk(result_paths[2]): 63 | if files: 64 | for f in files: 65 | path = os.path.join(root, f) 66 | highconfimgs = cv2.imread(path) # RGB image 67 | highconfimgs = cv2.cvtColor(highconfimgs, cv2.COLOR_BGR2RGB) 68 | h_r.append(highconfimgs) 69 | 70 | b_imgs = np.array(b_imgs) 71 | b_r = np.array(b_r) 72 | low_imgs = np.array(low_imgs) 73 | l_r = np.array(l_r) 74 | high_imgs = np.array(high_imgs) 75 | h_r = np.array(h_r) 76 | print(b_imgs.shape, b_r.shape, low_imgs.shape, l_r.shape, high_imgs.shape, h_r.shape) 77 | results = np.stack((b_imgs, b_r, low_imgs, l_r, high_imgs, h_r)) 78 | 79 | fig = plt.figure(figsize=(10, 10)) 80 | gs = gridspec.GridSpec(6, 10, wspace=0.1, hspace=0.1) 81 | 82 | for i in range(6): 83 | for j in range(10): 84 | ax = fig.add_subplot(gs[i, j]) 85 | ax.imshow(results[i, j], interpolation='none') 86 | ax.set_xticks([]) 87 | ax.set_yticks([]) 88 | gs.tight_layout(fig) 89 | plt.show() 90 | 91 | if __name__ == '__main__': 92 | plot(PATHS, RESULT_PATH) 93 | -------------------------------------------------------------------------------- /l0_yolov3.py: -------------------------------------------------------------------------------- 1 | import os 2 | 3 | # supress tensorflow logging other than errors 4 | # os.environ['TF_CPP_MIN_LOG_LEVEL'] = '1' 5 | import sys 6 | 7 | from keras import backend as K 8 | import numpy as np 9 | import random as rd 10 | import tensorflow as tf 11 | from tensorflow.python import debug as tf_debug 12 | from keras.models import Model 13 | from keras import losses 14 | 15 | from keras.layers.convolutional import Convolution2D, MaxPooling2D, ZeroPadding2D 16 | from keras.layers import Conv2D, MaxPooling2D, Input 17 | from keras.layers import Dense, Dropout, Activation, Flatten 18 | 19 | from keras.datasets import cifar10 20 | from keras.models import load_model 21 | from keras.callbacks import EarlyStopping 22 | 23 | from model.yolo_model import YOLO 24 | import cv2 25 | import matplotlib.pyplot as plt 26 | import time 27 | from tensorflow.python import debug as tf_debug 28 | from skimage import io 29 | 30 | YOLO_OUT_SHAPE = (13, 13, 3, 85) # yolo output shape 31 | IMAGE_SHAPE = (1, 416, 416, 3) # input image shape 32 | 33 | MAX_ITERATIONS = 10000 # number of iterations to perform gradient descent 34 | ABORT_EARLY = True # abort gradient descent upon first valid solution 35 | LEARNING_RATE = 1e-2 # larger values converge faster to less accurate results 36 | 37 | lower_bound = 0 38 | INITIAL_CONST = 1e2 39 | LARGEST_CONST = 1e10 40 | REDUCE_CONST = True # try to lower c each iteration; faster to set to false 41 | CONST_FACTOR = 2.0 # f>1, rate at which we increase constant, smaller better 42 | EXAMPLE_NUM = 10 43 | 44 | # ============runing setting================== 45 | CONFIDENCE = 0.3 46 | MAX_SEARCH = 5 47 | START_FROM = 0 48 | CUDA_GPU = '2' 49 | 50 | os.environ['CUDA_DEVICE_ORDER'] = 'PCI_BUS_ID' 51 | os.environ['CUDA_VISIBLE_DEVICES'] = CUDA_GPU 52 | 53 | PATH = 'adv_examples/L0/f3_eval/test/{0} confidence'.format(CONFIDENCE) 54 | 55 | def get_classes(file): 56 | """Get classes name. 57 | 58 | # Argument: 59 | file: classes name for database. 60 | 61 | # Returns 62 | class_names: List, classes name. 63 | 64 | """ 65 | with open(file) as f: 66 | class_names = f.readlines() 67 | class_names = [c.strip() for c in class_names] 68 | 69 | return class_names 70 | 71 | file = 'data/coco_classes.txt' 72 | all_classes = get_classes(file) 73 | 74 | def process_image(img): 75 | """ 76 | Resize, reduce and expand image. 77 | # Argument: 78 | img: original image. 79 | 80 | # Returns 81 | image: ndarray(64, 64, 3), processed image. 82 | """ 83 | image = cv2.resize(img, (416, 416), interpolation=cv2.INTER_CUBIC) 84 | image = np.array(image, dtype='float32') 85 | image /= 255. 86 | image = np.expand_dims(image, axis=0) 87 | return image 88 | 89 | 90 | def process_yolo_output(out, anchors, mask): 91 | """ 92 | Tensor op: Process output features. 93 | # Arguments 94 | out - tensor (?, N, N, 3, 4 + 1 +80), output feature map of yolo. 95 | anchors - List, anchors for box. 96 | mask - List, mask for anchors. 97 | 98 | # Returns 99 | boxes - tensor (N, N, 3, 4), x,y,w,h for per box. 100 | box_confidence - tensor (N, N, 3, 1), confidence for per box. 101 | box_class_probs - tensor (N, N, 3, 80), class probs for per box. 102 | """ 103 | grid_h, grid_w, num_boxes = map(int, out.shape[1: 4]) 104 | 105 | anchors = [anchors[i] for i in mask] 106 | # Reshape to batch, height, width, num_anchors, box_params. 107 | anchors_tensor = tf.reshape(tf.constant(anchors, dtype=tf.float32, name='anchor_tensor'), [1, 1, len(anchors), 2]) 108 | out = out[0] 109 | box_xy = tf.sigmoid(out[:, :, :, 0:2], name='box_xy') 110 | box_wh = tf.identity(tf.exp(out[:, :, :, 2:4]) * anchors_tensor, name='box_wh') 111 | 112 | box_confidence = tf.sigmoid(out[:, :, :, 4:5], name='objectness') 113 | box_class_probs = tf.sigmoid(out[:, :, :, 5:], name='class_probs') 114 | 115 | col = np.tile(np.arange(0, grid_w), grid_w).reshape(-1, grid_w) 116 | row = np.tile(np.arange(0, grid_h).reshape(-1, 1), grid_h) 117 | 118 | col = col.reshape(grid_h, grid_w, 1, 1).repeat(3, axis=-2) 119 | row = row.reshape(grid_h, grid_w, 1, 1).repeat(3, axis=-2) 120 | grid = tf.constant(np.concatenate((col, row), axis=-1), dtype=tf.float32) 121 | 122 | box_xy += grid 123 | box_xy /= (grid_w, grid_h) 124 | box_wh /= (416, 416) 125 | box_xy -= (box_wh / 2.) 126 | 127 | # boxes -> (13, 13, 3, 4) 128 | boxes = tf.concat([box_xy, box_wh], axis=-1) 129 | # box_confidence -> (13, 13, 3, 1) 26 52 130 | # box_class_probs -> (13, 13, 3, 80) 131 | boxes = tf.reshape(boxes, [int(boxes.shape[0]) ** 132 | 2, boxes.shape[2], boxes.shape[3]]) 133 | box_confidence = tf.reshape(box_confidence, 134 | [int(box_confidence.shape[0]) ** 2, 135 | box_confidence.shape[-2], 136 | box_confidence.shape[-1]]) 137 | box_class_probs = tf.reshape(box_class_probs, 138 | [int(box_class_probs.shape[0]) ** 2, 139 | box_class_probs.shape[-2], 140 | box_class_probs.shape[-1]], 141 | name='class_probs') 142 | return boxes, box_confidence, box_class_probs 143 | 144 | 145 | def process_output(raw_outs): 146 | """ 147 | Tensor op: Extract b, c, and s from raw outputs. 148 | # Args: 149 | raw_outs - Yolo raw output tensor. 150 | # 151 | boxes - Tensors. (N1**2+N2**2+N3**2, 3, 4), classes: (N1**2+N2**2+N3**2, 3, 1), scores: (N1**2+N2**2+N3**2, 3, 80) 152 | """ 153 | masks = [[6, 7, 8], [3, 4, 5], [0, 1, 2]] 154 | anchors = [[10, 13], [16, 30], [33, 23], [30, 61], [62, 45], [59, 119], [116, 90], [156, 198], [373, 326]] 155 | boxes, objecness, scores = [], [], [] 156 | 157 | for out, mask in zip(raw_outs, masks): 158 | # out -> (1, 13, 13, 3, 85) 159 | # mask -> [6, 7, 8] 160 | # boxes(13X13, 3, 4), box_confidence(13X13, 3, 1), 161 | # box_class_probs(13X13, 3, 80) | 26 X 26 | 162 | b, c, s = process_yolo_output(out, anchors, mask) 163 | if boxes == []: 164 | boxes = b 165 | objecness = c 166 | scores = s 167 | else: 168 | boxes = tf.concat([boxes, b], 0, name='xywh') 169 | objecness = tf.concat([objecness, c], 0, name='objectness') 170 | scores = tf.concat([scores, s], 0, name='class_probs') 171 | return boxes, objecness, scores 172 | 173 | 174 | def pdist(xy): 175 | """ 176 | Tensor op: Computes pairwise distance between each pair of points 177 | # Args: 178 | xy - [N,2] matrix representing N box position coordinates 179 | # Content: 180 | dists - [N,N] matrix of (squared) Euclidean distances 181 | # Return: 182 | expectation of the Euclidean distances 183 | """ 184 | xy2 = tf.reduce_sum(xy * xy, 1, True) 185 | dists = xy2 - 2 * tf.matmul(xy, tf.transpose(xy)) + tf.transpose(xy2) 186 | return tf.reduce_sum(dists) 187 | 188 | def output_to_pdist(bx, by): 189 | """ 190 | Tensor op: calculate expectation of box distance given yolo outpput bx & by. 191 | # Args: 192 | bx - YOLOv3 output batch x coordinates in shape (N, 3549, 3, 1) 193 | by - YOLOv3 output batch y coordinates in shape (N, 3549, 3, 1) 194 | """ 195 | bxby = tf.concat([bx, by], axis=-1) 196 | bxby = tf.reshape(bxby, [-1, 2]) 197 | return pdist(bxby) 198 | 199 | def pairwise_IoUs(bs1, bs2): 200 | """ 201 | Tensor op: Calculate pairwise IoUs given two sets of boxes. 202 | # Arguments: 203 | bs1, bs2 - tensor of boxes in shape (?, 4) 204 | # Content: 205 | X11,y11------x12,y11 X21,y21------x22,y21 206 | | | | | 207 | | | | | 208 | x11,y12-------x12,y12 x21,y22-------x22,y22 209 | # Returns: 210 | iou - a tensor of the matrix containing pairwise IoUs, in shape (?, ?) 211 | """ 212 | x11, y11, w1, h1 = tf.split(bs1, 4, axis=1) # (N, 1) 213 | x21, y21, w2, h2 = tf.split(bs2, 4, axis=1) # (N, 1) 214 | x12 = x11 + w1 215 | y12 = y11 + h1 216 | x22 = x21 + w2 217 | y22 = y21 + h2 218 | xA = tf.maximum(x11, tf.transpose(x21)) 219 | yA = tf.maximum(y11, tf.transpose(y21)) 220 | xB = tf.minimum(x12, tf.transpose(x22)) 221 | yB = tf.minimum(y12, tf.transpose(y22)) 222 | # prevent 0 area 223 | interArea = tf.maximum((xB - xA + 1), 0) * tf.maximum((yB - yA + 1), 0) 224 | 225 | boxAArea = (x12 - x11 + 1) * (y12 - y11 + 1) 226 | boxBArea = (x22 - x21 + 1) * (y22 - y21 + 1) 227 | 228 | iou = interArea / (boxAArea + tf.transpose(boxBArea) - interArea) 229 | return iou 230 | 231 | 232 | def expectation_of_IoUs(boxes): 233 | """ 234 | Tensor op: Calculate the expectation given all pairwise IoUs. 235 | # Arguments 236 | boxes - boxes of objects. It takes (?, 4) shaped tensor; 237 | # Returns 238 | expt - expectation of IoUs of box pairs. Scalar tensor. 239 | """ 240 | IoUs = pairwise_IoUs(boxes, boxes) 241 | expt = tf.reduce_mean(IoUs) 242 | return expt 243 | 244 | 245 | def expectation_of_IoUs_accross_classes(boxes, box_scores): 246 | """ 247 | Tensor op: Calculate IoU expectation for IoU expectations from different class. 248 | Arguments: 249 | #boxes - (3549, 3, 4) tensor output from yolo net 250 | #box_scores - (N1**2+N2**2+N3**2, 3, 80) tensor 251 | Content: 252 | #box_classes - (N1**2+N2**2+N3**2, 3, 1) tensor 253 | Returns: 254 | #expt_over_all_classes - The IoU expectation of box pairs over all classes. 255 | """ 256 | box_classes = tf.cast(tf.argmax(box_scores, axis=-1), tf.int32, name='box_classes') 257 | class_counts = tf.bincount(box_classes) 258 | dominating_cls = tf.argmax(class_counts) 259 | dominating_cls = tf.cast(dominating_cls, tf.int32) 260 | index = tf.equal(box_classes, dominating_cls) 261 | index = tf.cast(index, tf.int32) 262 | others, dominating_boxes = tf.dynamic_partition(boxes, index, num_partitions=2, name='dynamic_partition') 263 | expt_over_all_classes = expectation_of_IoUs(dominating_boxes) 264 | 265 | return expt_over_all_classes 266 | 267 | 268 | class YoloAttacker: 269 | """ 270 | Daedalus adversarial example generator based on the Yolo v3 model. 271 | """ 272 | 273 | def __init__(self, sess, model, learning_rate=LEARNING_RATE, 274 | max_iterations=MAX_ITERATIONS, abort_early=ABORT_EARLY, 275 | initial_const=INITIAL_CONST, largest_const=LARGEST_CONST, 276 | reduce_const=REDUCE_CONST, const_factor=CONST_FACTOR, 277 | independent_channels=False, lower_bound=lower_bound, max_search = MAX_SEARCH): 278 | 279 | self.model = model 280 | self.sess = sess 281 | 282 | self.LEARNING_RATE = learning_rate 283 | self.MAX_ITERATIONS = max_iterations 284 | self.ABORT_EARLY = abort_early 285 | self.INITIAL_CONST = initial_const 286 | self.LARGEST_CONST = largest_const 287 | self.REDUCE_CONST = reduce_const 288 | self.const_factor = const_factor 289 | self.independent_channels = independent_channels 290 | 291 | self.grad = self.gradient_descent(sess, model) 292 | 293 | self.confidence = CONFIDENCE 294 | self.lower_bound = lower_bound 295 | self.max_search = max_search 296 | 297 | self.search_iteration = 1 298 | 299 | def gradient_descent(self, sess, model): 300 | 301 | shape = IMAGE_SHAPE 302 | 303 | # the variable to optimize over 304 | modifier = tf.Variable(np.zeros(shape, dtype=np.float32)) 305 | 306 | # the variables we're going to hold, use for efficiency 307 | canchange = tf.Variable(np.zeros(shape), dtype=np.float32) 308 | simg = tf.Variable(np.zeros(shape, dtype=np.float32)) 309 | original = tf.Variable(np.zeros(shape, dtype=np.float32)) 310 | timg = tf.Variable(np.zeros(shape, dtype=np.float32)) 311 | const = tf.placeholder(tf.float32, []) 312 | 313 | # and the assignment to set the variables 314 | assign_modifier = tf.placeholder(np.float32, shape) 315 | assign_canchange = tf.placeholder(np.float32, shape) 316 | assign_simg = tf.placeholder(np.float32, shape) 317 | assign_original = tf.placeholder(np.float32, shape) 318 | assign_timg = tf.placeholder(np.float32, shape) 319 | 320 | # these are the variables to initialize when we run 321 | set_modifier = tf.assign(modifier, assign_modifier) 322 | setup = [] 323 | setup.append(tf.assign(canchange, assign_canchange)) 324 | setup.append(tf.assign(timg, assign_timg)) 325 | setup.append(tf.assign(original, assign_original)) 326 | setup.append(tf.assign(simg, assign_simg)) 327 | 328 | newimg = ((tf.tanh(modifier + simg) + 1) / 2) * canchange + (1 - canchange) * original 329 | 330 | self.outs = self.model._yolo(newimg) 331 | # [(1, 13, 13, 3, 85), (1, 26, 26, 3, 85), (1, 52, 52, 3, 85)] 332 | # (3549, 3, 4), (3549, 3, 1), (3549, 3, 80) | 13 X 13 + 26 X 26 + 52 X 52 333 | boxes, objectness, classprobs = process_output(self.outs) 334 | 335 | Iou_expt = expectation_of_IoUs_accross_classes(boxes, classprobs) 336 | self.bx = boxes[..., 0:1] 337 | self.by = boxes[..., 1:2] 338 | self.bw = boxes[..., 2:3] 339 | self.bh = boxes[..., 3:4] 340 | self.obj_scores = objectness 341 | self.class_probs = classprobs 342 | self.box_scores = tf.multiply(self.obj_scores, tf.reduce_max(self.class_probs, axis=-1, keepdims=True)) 343 | 344 | # # Optimisation metrics: 345 | self.l2dist = tf.reduce_sum(tf.square(newimg - (tf.tanh(timg) + 1) / 2), [1, 2, 3]) 346 | self.image_sum = tf.reduce_sum(newimg) 347 | 348 | # Define DDoS losses: loss must be a tensor here! 349 | # Make the objectness of all detections to be 1. 350 | self.loss1_1_x = tf.reduce_mean(tf.square(self.box_scores - 1), [-3, -2, -1]) # X 351 | 352 | # Minimising the size of all bounding box. 353 | self.f1 = tf.reduce_mean(Iou_expt) 354 | self.f2 = tf.reduce_mean(tf.square(tf.multiply(self.bw, self.bh)), [-3, -2, -1]) # a 355 | self.f3 = self.f2 + 1/output_to_pdist(self.bx, self.by) 356 | 357 | # add two loss terms together 358 | self.loss_adv = self.loss1_1_x + self.f2 359 | loss1 = tf.reduce_mean(const * self.loss_adv) 360 | loss2 = tf.reduce_mean(self.l2dist) 361 | loss = loss1 + loss2 362 | 363 | outgrad = tf.gradients(loss, [modifier])[0] 364 | 365 | # setup the adam optimizer and keep track of variables we're creating 366 | start_vars = set(x.name for x in tf.global_variables()) 367 | optimizer = tf.train.AdamOptimizer(self.LEARNING_RATE) 368 | train = optimizer.minimize(loss, var_list=[modifier]) 369 | 370 | end_vars = tf.global_variables() 371 | new_vars = [x for x in end_vars if x.name not in start_vars] 372 | init = tf.variables_initializer(var_list=[modifier, canchange, simg, 373 | original, timg] + new_vars) 374 | 375 | 376 | def doit(oimgs, starts, valid, CONST): 377 | # convert to tanh-space 378 | imgs = np.arctanh((np.array(oimgs) * 2 - 1) * .999999) 379 | starts = np.arctanh((np.array(starts) * 2 - 1) * .999999) 380 | 381 | # initialize the variables 382 | sess.run(init) 383 | sess.run(setup, {assign_timg: imgs, 384 | assign_simg: starts, 385 | assign_original: oimgs, 386 | assign_canchange: valid}) 387 | 388 | while self.search_iteration <= self.max_search: 389 | # try solving for each value of the constant 390 | print('=== try const ===', CONST, "|=== search_iteration ===", self.search_iteration) 391 | first_flag = True 392 | init_adv_losses = None 393 | for step in range(self.MAX_ITERATIONS): 394 | feed_dict = {const: CONST} 395 | 396 | # remember the old value 397 | oldmodifier = self.sess.run(modifier) 398 | 399 | # perform the update step 400 | _, works, l1= sess.run([train, loss1, self.loss_adv], feed_dict=feed_dict) 401 | if first_flag: 402 | init_adv_losses = l1 403 | first_flag = False 404 | 405 | def check_success(loss, init_loss): 406 | """ 407 | Check if the initial loss value has been reduced by 'self.confidence' percent 408 | """ 409 | return loss <= init_loss * (1 - self.confidence) 410 | 411 | if check_success(l1, init_adv_losses) and (self.ABORT_EARLY or step == CONST - 1): 412 | loss_shown, l2s, newimg_shown, l1 = sess.run([loss, loss2, newimg, self.loss_adv], feed_dict=feed_dict) 413 | l0_attack_pixel = np.sum(valid) 414 | # it worked previously, restore the old value and finish 415 | self.sess.run(set_modifier, {assign_modifier: oldmodifier}) 416 | grads, scores, nimg = sess.run((outgrad, self.outs, newimg), 417 | feed_dict=feed_dict) 418 | l2s = np.array([l2s]) 419 | return grads, scores, nimg, CONST, l2s 420 | 421 | self.lower_bound = max(self.lower_bound, CONST) 422 | if self.LARGEST_CONST < 1e9: 423 | CONST = (self.lower_bound + self.LARGEST_CONST) / 2 424 | else: 425 | CONST *= 10 426 | self.search_iteration += 1 427 | 428 | return doit 429 | 430 | def attack_single(self, img): 431 | """ 432 | Run the attack on a single image and label 433 | """ 434 | 435 | # the pixels we can change 436 | valid = np.ones((1, IMAGE_SHAPE[1], IMAGE_SHAPE[2], IMAGE_SHAPE[3])) 437 | 438 | # the previous image 439 | prev = np.copy(img).reshape((1, IMAGE_SHAPE[1], IMAGE_SHAPE[2], 440 | IMAGE_SHAPE[3])) 441 | last_solution = np.zeros((1,416,416,3)) 442 | last_distortion = np.zeros((1,)) 443 | last_const = np.zeros((1,)) 444 | const = self.INITIAL_CONST 445 | self.search_iteration = 1 446 | while True: 447 | # try to solve given this valid map 448 | res = self.grad([np.copy(img)], np.copy(prev), 449 | valid, const) 450 | if res == None: 451 | # the attack failed, we return this as our final answer 452 | print("the attack failed, we return this as our final answer") 453 | return last_solution, last_distortion, last_const 454 | 455 | # the attack succeeded, now we pick new pixels to set to 0 456 | restarted = False 457 | gradientnorm, scores, nimg, const, l2s = res 458 | 459 | # save the results 460 | last_solution = prev = nimg 461 | last_distortion = l2s 462 | last_const = np.array([const]) 463 | 464 | # adjust the value of const 465 | if self.REDUCE_CONST: 466 | self.search_iteration += 1 467 | self.LARGEST_CONST = min(self.LARGEST_CONST, const) 468 | if self.LARGEST_CONST < 1e9: 469 | const = (self.lower_bound + self.LARGEST_CONST) / 2 470 | print('*** calculate equal_count ***') 471 | equal_count = 416 ** 2 - np.sum(np.all(np.abs(img - nimg[0]) < .0001, axis=2)) 472 | print("Forced equal:", np.sum(1 - valid), 473 | "Equal count:", equal_count) 474 | if np.sum(valid) == 0: 475 | # if no pixels changed, return 476 | return [img], l2s, last_const 477 | 478 | if self.independent_channels: 479 | # we are allowed to change each channel independently 480 | valid = valid.flatten() 481 | totalchange = abs(nimg[0] - img) * np.abs(gradientnorm[0]) 482 | else: 483 | # we care only about which pixels change, not channels independently 484 | # compute total change as sum of change for each channel 485 | valid = valid.reshape((IMAGE_SHAPE[1] ** 2, IMAGE_SHAPE[3])) 486 | totalchange = abs(np.sum(nimg[0] - img, axis=2)) * np.sum(np.abs(gradientnorm[0]), axis=2) 487 | totalchange = totalchange.flatten() 488 | 489 | # set some of the pixels to 0 depending on their total change 490 | did = 0 491 | for e in np.argsort(totalchange): 492 | if np.all(valid[e]): 493 | did += 1 494 | valid[e] = 0 495 | 496 | if totalchange[e] > .01: 497 | # if this pixel changed a lot, skip 498 | break 499 | if did >= .3 * equal_count ** .5: 500 | # if we changed too many pixels, skip 501 | print('we changed too many pixels, skip') 502 | break 503 | 504 | valid = np.reshape(valid, (1, IMAGE_SHAPE[1], IMAGE_SHAPE[1], -1)) 505 | # total nums of be masked based on l2 result 506 | print("Now forced equal:", np.sum(1 - valid)) 507 | 508 | 509 | def attack(self, imgs): 510 | """ 511 | Perform the L_0 attack on the given images. 512 | """ 513 | r1 = [] 514 | r2 = [] 515 | for i, img in enumerate(imgs): 516 | print("Attack iteration", i) 517 | X_adv, dists, consts = self.attack_single(img) 518 | 519 | if not os.path.exists(PATH): 520 | os.makedirs(PATH) 521 | np.save(PATH + '/Distortions of image {0}.npy'.format(i), dists) 522 | for j in range(len(X_adv)): 523 | io.imsave(PATH + '/Best example of {1} CONST {0}.png'.format(consts, i+j), X_adv[j]) 524 | print('====== save the result:', path+'/Best example of {1} CONST {0}.png'.format(consts, i+j), '======') 525 | r1.extend(X_adv) 526 | r2.extend(dists) 527 | 528 | return np.array(r1), np.array(r2) 529 | 530 | if __name__ == '__main__': 531 | sess = tf.InteractiveSession() 532 | init = tf.global_variables_initializer() 533 | sess.run(init) 534 | 535 | ORACLE = YOLO(0.6, 0.5) # The auguments do not matter. 536 | 537 | X_test = [] 538 | i=0 539 | for (root, dirs, files) in os.walk('../COCO/val2017/'): 540 | if files: 541 | for f in files: 542 | # select 10 images 543 | if i >= EXAMPLE_NUM: 544 | break 545 | print(f) 546 | path = os.path.join(root, f) 547 | image = cv2.imread(path) 548 | image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB) # RGB 549 | image = process_image(image) 550 | X_test.append(image) 551 | i=i+1 552 | X_test = np.concatenate(X_test, axis=0) 553 | attacker = YoloAttacker(sess, ORACLE) 554 | attacker.attack(X_test) -------------------------------------------------------------------------------- /l2_ensemble.py: -------------------------------------------------------------------------------- 1 | import os 2 | import sys 3 | # supress tensorflow logging other than errors 4 | os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3' 5 | sys.path.insert(0, '..') 6 | 7 | from keras import backend as K 8 | import numpy as np 9 | import random as rd 10 | import tensorflow as tf 11 | from tensorflow.python import debug as tf_debug 12 | from keras.models import Model 13 | from keras import losses 14 | 15 | from keras.layers.convolutional import Convolution2D, MaxPooling2D, ZeroPadding2D 16 | from keras.layers import Conv2D, MaxPooling2D, Input 17 | from keras.layers import Dense, Dropout, Activation, Flatten 18 | 19 | from keras.models import load_model 20 | from keras.callbacks import EarlyStopping 21 | 22 | from keras_retinanet import models as retina_models 23 | from YOLOv3.model.yolo_model import YOLO 24 | import cv2 25 | import matplotlib.pyplot as plt 26 | from skimage import io 27 | import time 28 | 29 | # Parameter settings: 30 | GPU_ID = 0 # which gpu to used 31 | ATTACK_MODE = 'all' # select attack mode from 'all', 'most', 'least' and 'single'; 32 | ATTACK_CLASS = None # select the class to attack in 'single' mode 33 | CONFIDENCE = 0.3 # the confidence of attack 34 | EXAMPLE_NUM = 10 # total number of adversarial example to generate. 35 | BATCH_SIZE = 1 # number of adversarial example generated in each batch 36 | 37 | BINARY_SEARCH_STEPS = 5 # number of times to adjust the constsant with binary search 38 | INITIAL_consts = 1e1 # the initial constsant c to pick as a first guess 39 | CLASS_NUM = 80 # 80 for COCO dataset 40 | MAX_ITERATIONS = 10000 # number of iterations to perform gradient descent 41 | ABORT_EARLY = True # if we stop improving, abort gradient descent early 42 | LEARNING_RATE = 1e-2 # larger values converge faster to less accurate results 43 | IMAGE_SHAPE = (416, 416, 3) # input image shape 44 | SAVE_PATH = 'adv_examples/L2/f3/{0}/'.format(ATTACK_MODE) 45 | # select GPU to use 46 | os.environ["CUDA_VISIBLE_DEVICES"] = '{0}'.format(GPU_ID) 47 | 48 | def load_yolov3(): 49 | return YOLO(0.5, 0.5) 50 | 51 | def load_yolo(): 52 | return YOLO(0.5, 0.5) 53 | 54 | def load_retinanet(): 55 | model_path = os.path.join('model', 'resnet50_coco_best_v2.1.0.h5') 56 | # load retinanet model 57 | oracle = retina_models.load_model(model_path, backbone_name='resnet50') 58 | oracle.layers.pop() 59 | oracle.outputs = [oracle.layers[-2].output, oracle.layers[-1].output] #remove nms from original model 60 | oracle.layers[-1].outbound_nodes = [] 61 | oracle.summary() 62 | return oracle 63 | 64 | def process_image(img): 65 | """ 66 | Resize, reduce and expand image. 67 | # Argument: 68 | img: original image. 69 | 70 | # Returns 71 | image: ndarray(64, 64, 3), processed image. 72 | """ 73 | image = cv2.resize(img, (416, 416), 74 | interpolation=cv2.INTER_CUBIC) 75 | image = np.array(image, dtype='float32') 76 | image /= 255. 77 | image = np.expand_dims(image, axis=0) 78 | return image 79 | 80 | 81 | def process_yolo_output(out, anchors, mask): 82 | """ 83 | Tensor op: Process output features. 84 | # Arguments 85 | out - tensor (N, S, S, 3, 4+1+80), output feature map of yolo. 86 | anchors - List, anchors for box. 87 | mask - List, mask for anchors. 88 | # Returns 89 | boxes - tensor (N, S, S, 3, 4), x,y,w,h for per box. 90 | box_confidence - tensor (N, S, S, 3, 1), confidence for per box. 91 | box_class_probs - tensor (N, S, S, 3, 80), class probs for per box. 92 | """ 93 | batchsize, grid_h, grid_w, num_boxes = map(int, out.shape[0:4]) 94 | 95 | box_confidence = tf.sigmoid(out[..., 4:5], name='objectness') # (N, S, S, 3, 1) 96 | box_class_probs = tf.sigmoid(out[..., 5:], name='class_probs') # (N, S, S, 3, 80) 97 | 98 | anchors = np.array([anchors[i] for i in mask]) # Dimension of the used three anchor boxes [[x,x], [x,x], [x,x]]. 99 | # duplicate to shape (batch, height, width, num_anchors, box_params). 100 | anchors = np.repeat(anchors[np.newaxis, :, :], grid_w, axis=0) # (S, 3, 2) 101 | anchors = np.repeat(anchors[np.newaxis, :, :, :], grid_h, axis=0) # (S, S, 3, 2) 102 | anchors = np.repeat(anchors[np.newaxis, :, :, :, :], batchsize, axis=0) # (N, S, S, 3, 2) 103 | anchors_tensors = tf.constant(anchors, dtype=tf.float32, name='anchor_tensors') 104 | 105 | box_xy = tf.sigmoid(out[..., 0:2], name='box_xy') # (N, S, S, 3, 2) 106 | box_wh = tf.identity(tf.exp(out[..., 2:4]) * anchors_tensors, name='box_wh') # (N, S, S, 3, 2) 107 | 108 | col = np.tile(np.arange(0, grid_w), grid_w).reshape(-1, grid_w) 109 | row = np.tile(np.arange(0, grid_h).reshape(-1, 1), grid_h) 110 | 111 | col = col.reshape(grid_h, grid_w, 1, 1).repeat(3, axis=-2) 112 | row = row.reshape(grid_h, grid_w, 1, 1).repeat(3, axis=-2) 113 | grid = np.concatenate((col, row), axis=-1) #(13, 13, 3, 2) 114 | grid_batch = np.repeat(grid[np.newaxis, :, :, :, :], batchsize, axis=0) 115 | box_xy += grid_batch 116 | box_xy /= (grid_w, grid_h) 117 | box_wh /= (416, 416) 118 | box_xy -= (box_wh / 2.) 119 | 120 | # boxes -> (N, S, S, 3, 4) 121 | boxes = tf.concat([box_xy, box_wh], axis=-1) 122 | boxes = tf.reshape(boxes, [batchsize, -1, boxes.shape[-2], boxes.shape[-1]], name='boxes') #(N, S*S, 3, 4) 123 | # box_confidence -> (N, S, S, 3, 1) or 26 or 52 124 | # box_class_probs -> (N, S, S, 3, 80) 125 | box_confidence = tf.reshape(box_confidence, [batchsize, 126 | -1, 127 | box_confidence.shape[-2], 128 | box_confidence.shape[-1]], name='box_confidence') 129 | box_class_probs = tf.reshape(box_class_probs, [batchsize, 130 | -1, 131 | box_class_probs.shape[-2], 132 | box_class_probs.shape[-1]], name='class_probs') 133 | return boxes, box_confidence, box_class_probs 134 | 135 | 136 | def process_output(raw_outs): 137 | """ 138 | Tensor op: Extract b, c, and s from raw outputs. 139 | # Args: 140 | raw_outs - Yolo raw output tensor list [(N, 13, 13, 3, 85), (N, 26, 26, 3, 85), (N, 26, 26, 3, 85)]. 141 | # Returns: 142 | boxes - Tensors. (N, 3549, 3, 4), classes: (N, 3549, 3, 1), scores: (N, 3549, 3, 80) 143 | """ 144 | masks = [[6, 7, 8], [3, 4, 5], [0, 1, 2]] 145 | anchors = [[10, 13], [16, 30], [33, 23], 146 | [30, 61], [62, 45], [59, 119], 147 | [116, 90], [156, 198], [373, 326]] 148 | boxes, objecness, scores = [], [], [] 149 | 150 | for out, mask in zip(raw_outs, masks): 151 | # out -> (N, 13, 13, 3, 85) 152 | # mask -> one of the masks 153 | # boxes (N, 13X13, 3, 4), box_confidence (N, 13X13, 3, 1) 154 | # box_class_probs (13X13, 3, 80) | 26 X 26 | 155 | b, c, s = process_yolo_output(out, anchors, mask) 156 | if boxes == []: 157 | boxes = b 158 | objecness = c 159 | scores = s 160 | else: 161 | boxes = tf.concat([boxes, b], 1, name='xywh') 162 | objecness = tf.concat([objecness, c], 1, name='objectness') 163 | scores = tf.concat([scores, s], 1, name='class_probs') 164 | return boxes, objecness, scores 165 | 166 | class Daedalus: 167 | """ 168 | Daedalus adversarial example generator based on the Yolo v3 model. 169 | """ 170 | def __init__(self, sess, models, target_class=ATTACK_CLASS, attack_mode=ATTACK_MODE, img_shape=IMAGE_SHAPE, 171 | batch_size=BATCH_SIZE, confidence=CONFIDENCE, learning_rate=LEARNING_RATE, binary_search_steps=BINARY_SEARCH_STEPS, 172 | max_iterations=MAX_ITERATIONS, abort_early=ABORT_EARLY, initial_consts=INITIAL_consts, boxmin=0, boxmax=1): 173 | 174 | # self.sess = tf_debug.LocalCLIDebugWrapperSession(sess) 175 | self.sess = sess 176 | self.LEARNING_RATE = learning_rate 177 | self.MAX_ITERATIONS = max_iterations 178 | self.BINARY_SEARCH_STEPS = binary_search_steps 179 | self.ABORT_EARLY = abort_early 180 | self.initial_consts = initial_consts 181 | self.batch_size = batch_size 182 | self.repeat = binary_search_steps >= 6 183 | self.yolo3 = models[0] 184 | self.yolo = models[1] 185 | self.retinanet = models[2] 186 | self.confidence = confidence 187 | self.img_dimension = img_shape[0] 188 | self.target_class = target_class 189 | self.attack_mode = attack_mode 190 | 191 | def select_class(target_class, boxes, objectness, box_scores, mode='all'): 192 | box_classes = tf.cast(tf.argmax(box_scores, axis=-1), tf.int32, name='box_classes') 193 | class_counts = tf.bincount(box_classes) 194 | print(class_counts) 195 | if mode == 'all': 196 | selected_boxes = tf.reshape(boxes, [BATCH_SIZE, -1, 4]) 197 | selected_scores = tf.reshape(box_scores, [BATCH_SIZE, -1, CLASS_NUM]) 198 | if objectness == None: 199 | return selected_boxes, None, selected_scores 200 | selected_objectness = tf.reshape(objectness, [BATCH_SIZE, -1, 1]) 201 | return selected_boxes, selected_objectness, selected_scores 202 | elif mode == 'most': 203 | selected_cls = tf.argmax(class_counts) 204 | elif mode == 'least': 205 | class_counts = tf.where(tf.equal(class_counts,0), int(1e6)*tf.ones_like(class_counts, dtype=tf.int32), class_counts) 206 | selected_cls = tf.argmin(class_counts) 207 | elif mode == 'single': 208 | file = 'data/coco_classes.txt' 209 | with open(file) as f: 210 | class_names = f.readlines() 211 | class_names = [c.strip() for c in class_names] 212 | selected_cls = class_names.index(target_class) 213 | selected_cls = tf.cast(selected_cls, tf.int32) 214 | index = tf.equal(box_classes, selected_cls) 215 | index = tf.cast(index, tf.int32) 216 | _, selected_boxes = tf.dynamic_partition(boxes, index, num_partitions=2, name='dynamic_partition') 217 | _, selected_scores = tf.dynamic_partition(box_scores, index, num_partitions=2, name='dynamic_partition') 218 | selected_boxes = tf.reshape(selected_boxes, [BATCH_SIZE, -1, 4]) 219 | selected_scores = tf.reshape(selected_scores, [BATCH_SIZE, -1, CLASS_NUM]) 220 | if objectness == None: 221 | return selected_boxes, None, selected_scores 222 | _, selected_objectness = tf.dynamic_partition(objectness, index, num_partitions=2, name='dynamic_partition') 223 | selected_objectness = tf.reshape(selected_objectness, [BATCH_SIZE, -1, 1]) 224 | return selected_boxes, selected_objectness, selected_scores 225 | 226 | def yolov3_cg(images): 227 | # Get prediction from the model: 228 | outs = self.yolo3._yolo(images) 229 | # [(N, 13, 13, 3, 85), (N, 26, 26, 3, 85), (N, 52, 52, 3, 85)] to (N, 3549, 3, 4), (N, 3549, 3, 1), (N, 3549, 3, 80) 230 | boxes, objectness, classprobs = process_output(outs) 231 | boxes, objectness, classprobs = select_class(self.target_class, boxes, objectness, classprobs, mode=self.attack_mode) 232 | print(boxes, objectness, classprobs) 233 | self.yolo3bx = boxes[..., 0:1] 234 | self.yolo3by = boxes[..., 1:2] 235 | self.yolo3bw = boxes[..., 2:3] 236 | self.yolo3bh = boxes[..., 3:4] 237 | self.yolo3obj_scores = objectness 238 | self.yolo3class_probs = classprobs 239 | self.yolo3box_scores = tf.multiply(self.yolo3obj_scores, tf.reduce_max(self.yolo3class_probs, axis=-1, keepdims=True)) 240 | return 241 | 242 | def retina_cg(images): 243 | caffe_imgs = images * 255. 244 | caffe_imgs = caffe_imgs[..., ::-1] 245 | caffe_offsets = np.concatenate([103.939*np.ones((batch_size, 416, 416, 1)), 246 | 116.779*np.ones((batch_size, 416, 416, 1)), 247 | 123.68*np.ones((batch_size, 416, 416, 1))], axis=-1) 248 | caffe_imgs = caffe_imgs - caffe_offsets 249 | boxes, classprobs = self.retinanet(images) 250 | boxes, _, classprobs = select_class(self.target_class, boxes, None, classprobs, mode=self.attack_mode) 251 | print(boxes, classprobs) 252 | self.retinax1 = boxes[..., 0:1]/self.img_dimension 253 | self.retinay1 = boxes[..., 1:2]/self.img_dimension 254 | self.retinax2 = boxes[..., 2:3]/self.img_dimension 255 | self.retinay2 = boxes[..., 3:4]/self.img_dimension 256 | self.retinabw = tf.math.abs(self.retinax2 - self.retinax1) 257 | self.retinabh = tf.math.abs(self.retinay1 - self.retinay2) 258 | self.retinaclass_probs = classprobs 259 | self.retinabox_scores = tf.reduce_max(self.retinaclass_probs, axis=-1, keepdims=True) 260 | return 261 | 262 | # the perturbation we're going to optimize: 263 | perturbations = tf.Variable(np.zeros((batch_size, 264 | img_shape[0], 265 | img_shape[1], 266 | img_shape[2])), dtype=tf.float32, name='perturbations') 267 | # tf variables to sending data to tf: 268 | self.timgs = tf.Variable(np.zeros((batch_size, 269 | img_shape[0], 270 | img_shape[1], 271 | img_shape[2])), dtype=tf.float32, name='self.timgs') 272 | self.consts = tf.Variable(np.zeros(batch_size), dtype=tf.float32, name='self.consts') 273 | 274 | # and here's what we use to assign them: 275 | self.assign_timgs = tf.placeholder(tf.float32, (batch_size, 276 | img_shape[0], 277 | img_shape[1], 278 | img_shape[2])) 279 | self.assign_consts = tf.placeholder(tf.float32, [batch_size]) 280 | 281 | # Tensor operation: the resulting image, tanh'd to keep bounded from 282 | # boxmin to boxmax: 283 | self.boxmul = (boxmax - boxmin) / 2. 284 | self.boxplus = (boxmin + boxmax) / 2. 285 | self.newimgs = tf.tanh(perturbations + self.timgs) * self.boxmul + self.boxplus 286 | yolov3_cg(self.newimgs) 287 | retina_cg(self.newimgs) 288 | 289 | # Optimisation metrics: 290 | self.l2dist = tf.reduce_sum(tf.square(self.newimgs - (tf.tanh(self.timgs) * self.boxmul + self.boxplus)), [1, 2, 3]) 291 | 292 | # Define DDoS losses: loss must be a tensor here! 293 | # Make the box confidence of all detections to be 1. 294 | self.loss1_1_x = tf.reduce_mean(tf.square(self.yolo3box_scores-1), [-2,-1]) + tf.reduce_mean(tf.square(self.retinabox_scores-1), [-2,-1]) 295 | 296 | # Minimising the size of all bounding box. 297 | self.f3 = tf.reduce_mean(tf.square(tf.multiply(self.yolo3bw, self.yolo3bh)), [-2, -1]) + 1e3*tf.reduce_mean(tf.square(tf.multiply(self.retinabw, self.retinabh)), [-2, -1]) 298 | 299 | # add two loss terms together 300 | self.loss_adv = self.loss1_1_x + self.f3 301 | self.loss1 = tf.reduce_mean(self.consts * self.loss_adv) 302 | self.loss2 = tf.reduce_mean(self.l2dist) 303 | self.loss = self.loss1 + self.loss2 304 | 305 | # Setup the adam optimizer and keep track of variables we're creating 306 | start_vars = set(x.name for x in tf.global_variables()) 307 | optimizer = tf.train.AdamOptimizer(self.LEARNING_RATE) 308 | self.train = optimizer.minimize(self.loss, var_list=[perturbations]) 309 | end_vars = tf.global_variables() 310 | new_vars = [x for x in end_vars if x.name not in start_vars] 311 | 312 | # these are the variables to initialize when we run 313 | self.setup = [] 314 | self.setup.append(self.timgs.assign(self.assign_timgs)) 315 | self.setup.append(self.consts.assign(self.assign_consts)) 316 | self.init = tf.variables_initializer(var_list=[perturbations] + new_vars) 317 | 318 | def attack_batch(self, imgs): 319 | """ 320 | Run the attack on a batch of images and labels. 321 | """ 322 | 323 | def check_success(loss, init_loss): 324 | """ 325 | Check if the initial loss value has been reduced by 'self.confidence' percent 326 | """ 327 | return loss <= init_loss * (1 - self.confidence) 328 | 329 | batch_size = self.batch_size 330 | 331 | # convert images to arctanh-space 332 | imgs = np.arctanh((imgs - self.boxplus) / self.boxmul * 0.999999) 333 | 334 | # set the lower and upper bounds of the constsant. 335 | lower_bound = np.zeros(batch_size) 336 | consts = np.ones(batch_size) * self.initial_consts 337 | upper_bound = np.ones(batch_size) * 1e10 338 | 339 | # store the best l2, score, and image attack 340 | o_bestl2 = [1e10] * batch_size 341 | o_bestloss = [1e10] * batch_size 342 | o_bestattack = [np.zeros(imgs[0].shape)] * batch_size 343 | 344 | for outer_step in range(self.BINARY_SEARCH_STEPS): 345 | # completely reset adam's internal state. 346 | self.sess.run(self.init) 347 | 348 | # take in the current data batch. 349 | batch = imgs[:batch_size] 350 | 351 | # cache the current best l2 and score. 352 | bestl2 = [1e10] * batch_size 353 | # bestconfidence = [-1]*batch_size 354 | bestloss = [1e10] * batch_size 355 | 356 | # The last iteration (if we run many steps) repeat the search once. 357 | if self.repeat and outer_step == self.BINARY_SEARCH_STEPS - 1: 358 | consts = upper_bound 359 | 360 | # set the variables so that we don't have to send them over again. 361 | self.sess.run(self.setup, {self.assign_timgs: batch, 362 | self.assign_consts: consts}) 363 | 364 | obj_grads = tf.gradients(self.loss1_1_x, self.newimgs) 365 | print('objectness gradients:', sess.run(obj_grads)) 366 | loss_grads = tf.gradients(self.f3, self.newimgs) 367 | print('loss gradients:', sess.run(loss_grads)) 368 | 369 | # start gradient descent attack 370 | print('adjust c to:', sess.run(self.consts)) 371 | init_loss = sess.run(self.loss) 372 | init_adv_losses = sess.run(self.loss_adv) 373 | prev = init_loss * 1.1 374 | for iteration in range(self.MAX_ITERATIONS): 375 | # perform the attack on a single example 376 | _, l, l2s, l1s, nimgs, c = self.sess.run([self.train, self.loss, self.l2dist, self.loss_adv, self.newimgs, self.consts]) 377 | # print out the losses every 10% 378 | if iteration % (self.MAX_ITERATIONS // 10) == 0: 379 | print('===iteration:', iteration, '===') 380 | print('attacked box number:', sess.run(self.yolo3bw).shape, sess.run(self.retinabw).shape) 381 | print('loss values of box confidence and dimension:', sess.run([self.loss1_1_x, self.f3])) 382 | print('adversarial losses:', l1s) 383 | print('distortions:', l2s) 384 | 385 | # check if we should abort search if we're getting nowhere. 386 | if self.ABORT_EARLY and iteration % (self.MAX_ITERATIONS // 10) == 0: 387 | if l > prev * .9999: 388 | break 389 | prev = l 390 | 391 | # update the best result found so far 392 | for e, (l1, l2, ii) in enumerate(zip(l1s, l2s, nimgs)): 393 | if l2 < bestl2[e] and check_success(l1, init_adv_losses[e]): 394 | bestl2[e] = l2 395 | bestloss[e] = l1 396 | if l2 < o_bestl2[e] and check_success(l1, init_adv_losses[e]): 397 | o_bestl2[e] = l2 398 | o_bestloss[e] = l1 399 | o_bestattack[e] = ii 400 | 401 | # adjust the constsant as needed 402 | for e in range(batch_size): 403 | if check_success(l1s[e], init_adv_losses[e]): 404 | # success, divide consts by two 405 | upper_bound[e] = min(upper_bound[e], consts[e]) 406 | if upper_bound[e] < 1e9: 407 | consts[e] = (lower_bound[e] + upper_bound[e]) / 2 408 | else: 409 | # failure, either multiply by 10 if no solution found yet 410 | # or do binary search with the known upper bound 411 | lower_bound[e] = max(lower_bound[e], consts[e]) 412 | if upper_bound[e] < 1e9: 413 | consts[e] = (lower_bound[e] + upper_bound[e]) / 2 414 | else: 415 | consts[e] *= 10 416 | # return the best solution found 417 | o_bestl2 = np.array(o_bestl2) 418 | return o_bestattack, o_bestl2 419 | 420 | 421 | def attack(self, imgs): 422 | """ 423 | Perform the L_2 attack on the given images for the given targets. 424 | If self.targeted is true, then the targets represents the target labels. 425 | If self.targeted is false, then targets are the original class labels. 426 | """ 427 | r = [] 428 | ds = [] 429 | print('go up to', len(imgs)) 430 | for i in range(0, len(imgs), self.batch_size): 431 | print('tick', i) 432 | X_adv, dists = self.attack_batch(imgs[i:i + self.batch_size]) 433 | path = SAVE_PATH+'ensemble/{0} confidence'.format(self.confidence) 434 | if not os.path.exists(path): 435 | os.makedirs(path) 436 | np.save(path+'/Distortions of images {0} to {1}.npy'.format(i, i+self.batch_size), dists) 437 | for j in range(len(X_adv)): 438 | io.imsave(path+'/Best example of {1} Distortion {2}.png'.format(self.confidence, i+j, dists[j]), X_adv[j]) 439 | r.extend(X_adv) 440 | ds.extend(dists) 441 | return np.array(r), np.array(ds) 442 | 443 | 444 | if __name__ == '__main__': 445 | sess = tf.InteractiveSession() 446 | init = tf.global_variables_initializer() 447 | sess.run(init) 448 | ORACLEs = [load_yolov3(), load_yolo(), load_retinanet()] # The auguments do not matter. 449 | attacker = Daedalus(sess, ORACLEs) 450 | 451 | X_test = [] 452 | for (root, dirs, files) in os.walk('../Datasets/COCO/val2017/'): 453 | if files: 454 | for f in files: 455 | print(f) 456 | path = os.path.join(root, f) 457 | image = cv2.imread(path) 458 | image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB) # RGB 459 | image = process_image(image) 460 | #image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR) 461 | X_test.append(image) 462 | EXAMPLE_NUM -= 1 463 | if EXAMPLE_NUM == 0: 464 | break 465 | X_test = np.concatenate(X_test, axis=0) 466 | 467 | start = time.time() 468 | X_adv, distortions = attacker.attack(X_test) 469 | end = time.time() 470 | print('time: {0:.2f}s'.format((end - start)*0.2)) 471 | f = open('f2 runtime.txt', 'a') 472 | f.write('time: {0:.2f}s\n'.format((end - start)*0.2)) 473 | f.close() 474 | np.savez(SAVE_PATH+'ensemble/{} confidence/Daedalus example batch.npz'.format(CONFIDENCE), X_adv=X_adv, distortions=distortions) 475 | writer = tf.summary.FileWriter("log", sess.graph) 476 | writer.close() 477 | -------------------------------------------------------------------------------- /l2_retinanet.py: -------------------------------------------------------------------------------- 1 | import os 2 | import sys 3 | # supress tensorflow logging other than errors 4 | os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3' 5 | sys.path.insert(0, '..') 6 | 7 | from keras import backend as K 8 | import numpy as np 9 | import random as rd 10 | import tensorflow as tf 11 | from tensorflow.python import debug as tf_debug 12 | from keras.models import Model 13 | from keras import losses 14 | 15 | from keras.layers.convolutional import Convolution2D, MaxPooling2D, ZeroPadding2D 16 | from keras.layers import Conv2D, MaxPooling2D, Input 17 | from keras.layers import Dense, Dropout, Activation, Flatten 18 | 19 | from keras.datasets import cifar10 20 | from keras.models import load_model 21 | from keras.callbacks import EarlyStopping 22 | 23 | from keras_retinanet import models 24 | import cv2 25 | import matplotlib.pyplot as plt 26 | from skimage import io 27 | import time 28 | 29 | # Parameter settings: 30 | GPU_ID = 0 # which gpu to used 31 | ATTACK_MODE = 'all' # select attack mode from 'all', 'most', 'least' and 'single'; 32 | ATTACK_CLASS = None # select the class to attack in 'single' mode 33 | CONFIDENCE = 0.3 # the confidence of attack 34 | EXAMPLE_NUM = 10 # total number of adversarial example to generate. 35 | BATCH_SIZE = 1 # number of adversarial example generated in each batch 36 | 37 | BINARY_SEARCH_STEPS = 5 # number of times to adjust the constsant with binary search 38 | INITIAL_consts = 50 # the initial constsant c to pick as a first guess 39 | CLASS_NUM = 80 # 80 for COCO dataset 40 | MAX_ITERATIONS = 10000 # number of iterations to perform gradient descent 41 | ABORT_EARLY = True # if we stop improving, abort gradient descent early 42 | LEARNING_RATE = 1e-2 # larger values converge faster to less accurate results 43 | IMAGE_SHAPE = (416, 416, 3) # input image shape 44 | SAVE_PATH = 'adv_examples/L2/f3/{0}/'.format(ATTACK_MODE) 45 | # select GPU to use 46 | os.environ["CUDA_VISIBLE_DEVICES"] = '{0}'.format(GPU_ID) 47 | 48 | 49 | def process_image(img): 50 | """ 51 | Resize, reduce and expand image. 52 | # Argument: 53 | img: original image. 54 | 55 | # Returns 56 | image: ndarray(64, 64, 3), processed image. 57 | """ 58 | image = cv2.resize(img, (416, 416), 59 | interpolation=cv2.INTER_CUBIC) 60 | image = np.array(image, dtype='float32') 61 | image /= 255. 62 | image = np.expand_dims(image, axis=0) 63 | return image 64 | 65 | class Daedalus: 66 | """ 67 | Daedalus adversarial example generator based on the Yolo v3 model. 68 | """ 69 | def __init__(self, sess, model, target_class=ATTACK_CLASS, attack_mode=ATTACK_MODE, img_shape=IMAGE_SHAPE, 70 | batch_size=BATCH_SIZE, confidence=CONFIDENCE, learning_rate=LEARNING_RATE, binary_search_steps=BINARY_SEARCH_STEPS, 71 | max_iterations=MAX_ITERATIONS, abort_early=ABORT_EARLY, initial_consts=INITIAL_consts, boxmin=0, boxmax=1): 72 | 73 | # self.sess = tf_debug.LocalCLIDebugWrapperSession(sess) 74 | self.sess = sess 75 | self.LEARNING_RATE = learning_rate 76 | self.MAX_ITERATIONS = max_iterations 77 | self.BINARY_SEARCH_STEPS = binary_search_steps 78 | self.ABORT_EARLY = abort_early 79 | self.initial_consts = initial_consts 80 | self.batch_size = batch_size 81 | self.repeat = binary_search_steps >= 6 82 | self.detection_model = model 83 | self.confidence = confidence 84 | self.img_dimension = img_shape[0] 85 | self.target_class = target_class 86 | self.attack_mode = attack_mode 87 | 88 | def select_class(target_class, boxes, objectness, box_scores, mode='all'): 89 | box_classes = tf.cast(tf.argmax(box_scores, axis=-1), tf.int32, name='box_classes') 90 | class_counts = tf.bincount(box_classes) 91 | print(class_counts) 92 | if mode == 'all': 93 | selected_boxes = tf.reshape(boxes, [BATCH_SIZE, -1, 4]) 94 | selected_scores = tf.reshape(box_scores, [BATCH_SIZE, -1, CLASS_NUM]) 95 | if objectness == None: 96 | return selected_boxes, None, selected_scores 97 | selected_objectness = tf.reshape(objectness, [BATCH_SIZE, -1, 1]) 98 | return selected_boxes, selected_objectness, selected_scores 99 | elif mode == 'most': 100 | selected_cls = tf.argmax(class_counts) 101 | elif mode == 'least': 102 | class_counts = tf.where(tf.equal(class_counts,0), int(1e6)*tf.ones_like(class_counts, dtype=tf.int32), class_counts) 103 | selected_cls = tf.argmin(class_counts) 104 | elif mode == 'single': 105 | file = 'data/coco_classes.txt' 106 | with open(file) as f: 107 | class_names = f.readlines() 108 | class_names = [c.strip() for c in class_names] 109 | selected_cls = class_names.index(target_class) 110 | selected_cls = tf.cast(selected_cls, tf.int32) 111 | index = tf.equal(box_classes, selected_cls) 112 | index = tf.cast(index, tf.int32) 113 | _, selected_boxes = tf.dynamic_partition(boxes, index, num_partitions=2, name='dynamic_partition') 114 | _, selected_scores = tf.dynamic_partition(box_scores, index, num_partitions=2, name='dynamic_partition') 115 | selected_boxes = tf.reshape(selected_boxes, [BATCH_SIZE, -1, 4]) 116 | selected_scores = tf.reshape(selected_scores, [BATCH_SIZE, -1, CLASS_NUM]) 117 | if objectness == None: 118 | return selected_boxes, None, selected_scores 119 | _, selected_objectness = tf.dynamic_partition(objectness, index, num_partitions=2, name='dynamic_partition') 120 | selected_objectness = tf.reshape(selected_objectness, [BATCH_SIZE, -1, 1]) 121 | return selected_boxes, selected_objectness, selected_scores 122 | 123 | # the perturbation we're going to optimize: 124 | perturbations = tf.Variable(np.zeros((batch_size, 125 | img_shape[0], 126 | img_shape[1], 127 | img_shape[2])), dtype=tf.float32, name='perturbations') 128 | # tf variables to sending data to tf: 129 | self.timgs = tf.Variable(np.zeros((batch_size, 130 | img_shape[0], 131 | img_shape[1], 132 | img_shape[2])), dtype=tf.float32, name='self.timgs') 133 | self.consts = tf.Variable(np.zeros(batch_size), dtype=tf.float32, name='self.consts') 134 | 135 | # and here's what we use to assign them: 136 | self.assign_timgs = tf.placeholder(tf.float32, (batch_size, 137 | img_shape[0], 138 | img_shape[1], 139 | img_shape[2])) 140 | self.assign_consts = tf.placeholder(tf.float32, [batch_size]) 141 | 142 | # Tensor operation: the resulting image, tanh'd to keep bounded from 143 | # boxmin to boxmax: 144 | self.boxmul = (boxmax - boxmin) / 2. 145 | self.boxplus = (boxmin + boxmax) / 2. 146 | self.newimgs = tf.tanh(perturbations + self.timgs) * self.boxmul + self.boxplus 147 | 148 | caffe_imgs = self.newimgs * 255. 149 | caffe_imgs = caffe_imgs[..., ::-1] 150 | caffe_offsets = np.concatenate([103.939*np.ones((batch_size, 416, 416, 1)), 151 | 116.779*np.ones((batch_size, 416, 416, 1)), 152 | 123.68*np.ones((batch_size, 416, 416, 1))], axis=-1) 153 | caffe_imgs = caffe_imgs - caffe_offsets 154 | 155 | # Get prediction from the model: 156 | boxes, classprobs = self.detection_model(caffe_imgs) 157 | boxes, _, classprobs = select_class(self.target_class, boxes, None, classprobs, mode=self.attack_mode) 158 | print(boxes, classprobs) 159 | self.x1 = boxes[..., 0:1]/self.img_dimension 160 | self.y1 = boxes[..., 1:2]/self.img_dimension 161 | self.x2 = boxes[..., 2:3]/self.img_dimension 162 | self.y2 = boxes[..., 3:4]/self.img_dimension 163 | self.bw = tf.math.abs(self.x2 - self.x1) 164 | self.bh = tf.math.abs(self.y1 - self.y2) 165 | self.class_probs = classprobs 166 | self.box_scores = tf.reduce_max(self.class_probs, axis=-1, keepdims=True) 167 | 168 | # Optimisation metrics: 169 | self.l2dist = tf.reduce_sum(tf.square(self.newimgs - (tf.tanh(self.timgs) * self.boxmul + self.boxplus)), [1, 2, 3]) 170 | 171 | # Define DDoS losses: loss must be a tensor here! 172 | # Make the box confidence of all detections to be 1. 173 | self.loss1_1_x = tf.reduce_mean(tf.square(self.box_scores - 1), [-2, -1]) 174 | 175 | # Minimising the size of all bounding box. 176 | self.f3 = 1e1 * tf.reduce_mean(tf.square(tf.multiply(self.bw, self.bh)), [-2, -1]) 177 | 178 | # add two loss terms together 179 | self.loss_adv = self.loss1_1_x + self.f3 180 | self.loss1 = tf.reduce_mean(self.consts * self.loss_adv) 181 | self.loss2 = tf.reduce_mean(self.l2dist) 182 | self.loss = self.loss1 + self.loss2 183 | 184 | # Setup the adam optimizer and keep track of variables we're creating 185 | start_vars = set(x.name for x in tf.global_variables()) 186 | optimizer = tf.train.AdamOptimizer(self.LEARNING_RATE) 187 | self.train = optimizer.minimize(self.loss, var_list=[perturbations]) 188 | end_vars = tf.global_variables() 189 | new_vars = [x for x in end_vars if x.name not in start_vars] 190 | 191 | # these are the variables to initialize when we run 192 | self.setup = [] 193 | self.setup.append(self.timgs.assign(self.assign_timgs)) 194 | self.setup.append(self.consts.assign(self.assign_consts)) 195 | self.init = tf.variables_initializer(var_list=[perturbations] + new_vars) 196 | 197 | def attack_batch(self, imgs): 198 | """ 199 | Run the attack on a batch of images and labels. 200 | """ 201 | 202 | def check_success(loss, init_loss): 203 | """ 204 | Check if the initial loss value has been reduced by 'self.confidence' percent 205 | """ 206 | return loss <= init_loss * (1 - self.confidence) 207 | 208 | batch_size = self.batch_size 209 | 210 | # convert images to arctanh-space 211 | imgs = np.arctanh((imgs - self.boxplus) / self.boxmul * 0.999999) 212 | 213 | # set the lower and upper bounds of the constsant. 214 | lower_bound = np.zeros(batch_size) 215 | consts = np.ones(batch_size) * self.initial_consts 216 | upper_bound = np.ones(batch_size) * 1e10 217 | 218 | # store the best l2, score, and image attack 219 | o_bestl2 = [1e10] * batch_size 220 | o_bestloss = [1e10] * batch_size 221 | o_bestattack = [np.zeros(imgs[0].shape)] * batch_size 222 | 223 | for outer_step in range(self.BINARY_SEARCH_STEPS): 224 | # completely reset adam's internal state. 225 | self.sess.run(self.init) 226 | 227 | # take in the current data batch. 228 | batch = imgs[:batch_size] 229 | 230 | # cache the current best l2 and score. 231 | bestl2 = [1e10] * batch_size 232 | # bestconfidence = [-1]*batch_size 233 | bestloss = [1e10] * batch_size 234 | 235 | # The last iteration (if we run many steps) repeat the search once. 236 | if self.repeat and outer_step == self.BINARY_SEARCH_STEPS - 1: 237 | consts = upper_bound 238 | 239 | # set the variables so that we don't have to send them over again. 240 | self.sess.run(self.setup, {self.assign_timgs: batch, 241 | self.assign_consts: consts}) 242 | 243 | # start gradient descent attack 244 | print('adjust c to:', sess.run(self.consts)) 245 | init_loss = sess.run(self.loss) 246 | init_adv_losses = sess.run(self.loss_adv) 247 | prev = init_loss * 1.1 248 | for iteration in range(self.MAX_ITERATIONS): 249 | # perform the attack on a single example 250 | _, l, l2s, l1s, nimgs, c = self.sess.run([self.train, self.loss, self.l2dist, self.loss_adv, self.newimgs, self.consts]) 251 | # print out the losses every 10% 252 | if iteration % (self.MAX_ITERATIONS // 10) == 0: 253 | print('===iteration:', iteration, '===') 254 | print('attacked box number:', sess.run(self.bw).shape) 255 | print('loss values of box confidence and dimension:', sess.run([self.loss1_1_x, self.f3])) 256 | print('adversarial losses:', l1s) 257 | print('distortions:', l2s) 258 | path = SAVE_PATH+'retinanet/{0} confidence'.format(self.confidence) 259 | if not os.path.exists(path): 260 | os.makedirs(path) 261 | #[io.imsave(path+'/debug_img_{0}Iteration_{1}.png'.format(i, iteration), nimgs[i]) for i in range(nimgs.shape[0])] 262 | 263 | # check if we should abort search if we're getting nowhere. 264 | if self.ABORT_EARLY and iteration % (self.MAX_ITERATIONS // 10) == 0: 265 | if l > prev * .9999: 266 | break 267 | prev = l 268 | 269 | # update the best result found so far 270 | for e, (l1, l2, ii) in enumerate(zip(l1s, l2s, nimgs)): 271 | if l2 < bestl2[e] and check_success(l1, init_adv_losses[e]): 272 | bestl2[e] = l2 273 | bestloss[e] = l1 274 | if l2 < o_bestl2[e] and check_success(l1, init_adv_losses[e]): 275 | o_bestl2[e] = l2 276 | o_bestloss[e] = l1 277 | o_bestattack[e] = ii 278 | 279 | # adjust the constsant as needed 280 | for e in range(batch_size): 281 | if check_success(l1s[e], init_adv_losses[e]): 282 | # success, divide consts by two 283 | upper_bound[e] = min(upper_bound[e], consts[e]) 284 | if upper_bound[e] < 1e9: 285 | consts[e] = (lower_bound[e] + upper_bound[e]) / 2 286 | else: 287 | # failure, either multiply by 10 if no solution found yet 288 | # or do binary search with the known upper bound 289 | lower_bound[e] = max(lower_bound[e], consts[e]) 290 | if upper_bound[e] < 1e9: 291 | consts[e] = (lower_bound[e] + upper_bound[e]) / 2 292 | else: 293 | consts[e] *= 10 294 | # return the best solution found 295 | o_bestl2 = np.array(o_bestl2) 296 | return o_bestattack, o_bestl2 297 | 298 | 299 | def attack(self, imgs): 300 | """ 301 | Perform the L_2 attack on the given images for the given targets. 302 | If self.targeted is true, then the targets represents the target labels. 303 | If self.targeted is false, then targets are the original class labels. 304 | """ 305 | r = [] 306 | ds = [] 307 | print('go up to', len(imgs)) 308 | for i in range(0, len(imgs), self.batch_size): 309 | print('tick', i) 310 | X_adv, dists = self.attack_batch(imgs[i:i + self.batch_size]) 311 | path = SAVE_PATH+'retinanet/{0} confidence'.format(self.confidence) 312 | if not os.path.exists(path): 313 | os.makedirs(path) 314 | np.save(path+'/Distortions of images {0} to {1}.npy'.format(i, i+self.batch_size), dists) 315 | for j in range(len(X_adv)): 316 | io.imsave(path+'/Best example of {1} Distortion {2}.png'.format(self.confidence, i+j, dists[j]), X_adv[j]) 317 | r.extend(X_adv) 318 | ds.extend(dists) 319 | return np.array(r), np.array(ds) 320 | 321 | 322 | if __name__ == '__main__': 323 | sess = tf.InteractiveSession() 324 | init = tf.global_variables_initializer() 325 | sess.run(init) 326 | # models can be downloaded here: https://github.com/fizyr/keras-retinanet/releases 327 | model_path = os.path.join('model', 'resnet50_coco_best_v2.1.0.h5') 328 | # load retinanet model 329 | ORACLE = models.load_model(model_path, backbone_name='resnet50') 330 | ORACLE.layers.pop() 331 | ORACLE.outputs = [ORACLE.layers[-2].output, ORACLE.layers[-1].output] #remove nms from original model 332 | ORACLE.layers[-1].outbound_nodes = [] 333 | ORACLE.summary() 334 | X_test = [] 335 | for (root, dirs, files) in os.walk('../Datasets/COCO/val2017'): 336 | if files: 337 | for f in files: 338 | print(f) 339 | path = os.path.join(root, f) 340 | image = cv2.imread(path) 341 | image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB) # RGB 342 | image = process_image(image) 343 | X_test.append(image) 344 | EXAMPLE_NUM -= 1 345 | if EXAMPLE_NUM == 0: 346 | break 347 | X_test = np.concatenate(X_test, axis=0) 348 | attacker = Daedalus(sess, ORACLE) 349 | X_adv, distortions = attacker.attack(X_test) 350 | np.savez(SAVE_PATH+'retinanet/{0} confidence/Daedalus example batch.npz'.format(CONFIDENCE), X_adv=X_adv, distortions=distortions) 351 | writer = tf.summary.FileWriter("log", sess.graph) 352 | writer.close() 353 | -------------------------------------------------------------------------------- /l2_yolov3.py: -------------------------------------------------------------------------------- 1 | import os 2 | import sys 3 | # supress tensorflow logging other than errors 4 | #os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3' 5 | 6 | from keras import backend as K 7 | import numpy as np 8 | import random as rd 9 | import tensorflow as tf 10 | from tensorflow.python import debug as tf_debug 11 | from keras.models import Model 12 | from keras import losses 13 | 14 | from keras.layers.convolutional import Convolution2D, MaxPooling2D, ZeroPadding2D 15 | from keras.layers import Conv2D, MaxPooling2D, Input 16 | from keras.layers import Dense, Dropout, Activation, Flatten 17 | 18 | from keras.models import load_model 19 | from keras.callbacks import EarlyStopping 20 | 21 | from model.yolo_model import YOLO 22 | import cv2 23 | import matplotlib.pyplot as plt 24 | from skimage import io 25 | import time 26 | 27 | # Parameter settings: 28 | GPU_ID = 0 # which gpu to used 29 | ATTACK_MODE = 'all' # select attack mode from 'all', 'most', 'least' and 'single'; 30 | ATTACK_CLASS = 'person' # select the class to attack in 'single' mode 31 | CONFIDENCE = 0.3 # the confidence of attack 32 | EXAMPLE_NUM = 10 # total number of adversarial example to generate. 33 | BATCH_SIZE = 1 # number of adversarial example generated in each batch 34 | 35 | BINARY_SEARCH_STEPS = 5 # number of times to adjust the constsant with binary search 36 | INITIAL_consts = 2 # the initial constsant c to pick as a first guess 37 | CLASS_NUM = 80 # 80 for COCO dataset 38 | MAX_ITERATIONS = 10000 # number of iterations to perform gradient descent 39 | ABORT_EARLY = True # if we stop improving, abort gradient descent early 40 | LEARNING_RATE = 1e-2 # larger values converge faster to less accurate results 41 | IMAGE_SHAPE = (416, 416, 3) # input image shape 42 | SAVE_PATH = 'adv_examples/L2/f3/{0}/'.format(ATTACK_MODE) 43 | # select GPU to use 44 | os.environ["CUDA_VISIBLE_DEVICES"] = '{0}'.format(GPU_ID) 45 | 46 | 47 | def process_image(img): 48 | """ 49 | Resize, reduce and expand image. 50 | # Argument: 51 | img: original image. 52 | 53 | # Returns 54 | image: ndarray(64, 64, 3), processed image. 55 | """ 56 | image = cv2.resize(img, (416, 416), 57 | interpolation=cv2.INTER_CUBIC) 58 | image = np.array(image, dtype='float32') 59 | image /= 255. 60 | image = np.expand_dims(image, axis=0) 61 | return image 62 | 63 | 64 | def process_yolo_output(out, anchors, mask): 65 | """ 66 | Tensor op: Process output features. 67 | # Arguments 68 | out - tensor (N, S, S, 3, 4+1+80), output feature map of yolo. 69 | anchors - List, anchors for box. 70 | mask - List, mask for anchors. 71 | # Returns 72 | boxes - tensor (N, S, S, 3, 4), x,y,w,h for per box. 73 | box_confidence - tensor (N, S, S, 3, 1), confidence for per box. 74 | box_class_probs - tensor (N, S, S, 3, 80), class probs for per box. 75 | """ 76 | batchsize, grid_h, grid_w, num_boxes = map(int, out.shape[0:4]) 77 | 78 | box_confidence = tf.sigmoid(out[..., 4:5], name='objectness') # (N, S, S, 3, 1) 79 | box_class_probs = tf.sigmoid(out[..., 5:], name='class_probs') # (N, S, S, 3, 80) 80 | 81 | anchors = np.array([anchors[i] for i in mask]) # Dimension of the used three anchor boxes [[x,x], [x,x], [x,x]]. 82 | # duplicate to shape (batch, height, width, num_anchors, box_params). 83 | anchors = np.repeat(anchors[np.newaxis, :, :], grid_w, axis=0) # (S, 3, 2) 84 | anchors = np.repeat(anchors[np.newaxis, :, :, :], grid_h, axis=0) # (S, S, 3, 2) 85 | anchors = np.repeat(anchors[np.newaxis, :, :, :, :], batchsize, axis=0) # (N, S, S, 3, 2) 86 | anchors_tensors = tf.constant(anchors, dtype=tf.float32, name='anchor_tensors') 87 | 88 | box_xy = tf.sigmoid(out[..., 0:2], name='box_xy') # (N, S, S, 3, 2) 89 | box_wh = tf.identity(tf.exp(out[..., 2:4]) * anchors_tensors, name='box_wh') # (N, S, S, 3, 2) 90 | 91 | col = np.tile(np.arange(0, grid_w), grid_w).reshape(-1, grid_w) 92 | row = np.tile(np.arange(0, grid_h).reshape(-1, 1), grid_h) 93 | 94 | col = col.reshape(grid_h, grid_w, 1, 1).repeat(3, axis=-2) 95 | row = row.reshape(grid_h, grid_w, 1, 1).repeat(3, axis=-2) 96 | grid = np.concatenate((col, row), axis=-1) #(13, 13, 3, 2) 97 | grid_batch = np.repeat(grid[np.newaxis, :, :, :, :], batchsize, axis=0) 98 | box_xy += grid_batch 99 | box_xy /= (grid_w, grid_h) 100 | box_wh /= (416, 416) 101 | box_xy -= (box_wh / 2.) 102 | 103 | # boxes -> (N, S, S, 3, 4) 104 | boxes = tf.concat([box_xy, box_wh], axis=-1) 105 | boxes = tf.reshape(boxes, [batchsize, -1, boxes.shape[-2], boxes.shape[-1]], name='boxes') #(N, S*S, 3, 4) 106 | # box_confidence -> (N, S, S, 3, 1) or 26 or 52 107 | # box_class_probs -> (N, S, S, 3, 80) 108 | box_confidence = tf.reshape(box_confidence, [batchsize, 109 | -1, 110 | box_confidence.shape[-2], 111 | box_confidence.shape[-1]], name='box_confidence') 112 | box_class_probs = tf.reshape(box_class_probs, [batchsize, 113 | -1, 114 | box_class_probs.shape[-2], 115 | box_class_probs.shape[-1]], name='class_probs') 116 | return boxes, box_confidence, box_class_probs 117 | 118 | 119 | def process_output(raw_outs): 120 | """ 121 | Tensor op: Extract b, c, and s from raw outputs. 122 | # Args: 123 | raw_outs - Yolo raw output tensor list [(N, 13, 13, 3, 85), (N, 26, 26, 3, 85), (N, 26, 26, 3, 85)]. 124 | # Returns: 125 | boxes - Tensors. (N, 3549, 3, 4), classes: (N, 3549, 3, 1), scores: (N, 3549, 3, 80) 126 | """ 127 | masks = [[6, 7, 8], [3, 4, 5], [0, 1, 2]] 128 | anchors = [[10, 13], [16, 30], [33, 23], 129 | [30, 61], [62, 45], [59, 119], 130 | [116, 90], [156, 198], [373, 326]] 131 | boxes, objecness, scores = [], [], [] 132 | 133 | for out, mask in zip(raw_outs, masks): 134 | # out -> (N, 13, 13, 3, 85) 135 | # mask -> one of the masks 136 | # boxes (N, 13X13, 3, 4), box_confidence (N, 13X13, 3, 1) 137 | # box_class_probs (13X13, 3, 80) | 26 X 26 | 138 | b, c, s = process_yolo_output(out, anchors, mask) 139 | if boxes == []: 140 | boxes = b 141 | objecness = c 142 | scores = s 143 | else: 144 | boxes = tf.concat([boxes, b], 1, name='xywh') 145 | objecness = tf.concat([objecness, c], 1, name='objectness') 146 | scores = tf.concat([scores, s], 1, name='class_probs') 147 | return boxes, objecness, scores 148 | 149 | 150 | def pdist(xy): 151 | """ 152 | Tensor op: Computes pairwise distance between each pair of points 153 | # Args: 154 | xy - [N,2] matrix representing N box position coordinates 155 | # Content: 156 | dists - [N,N] matrix of (squared) Euclidean distances 157 | # Return: 158 | expectation of the Euclidean distances 159 | """ 160 | xy2 = tf.reduce_sum(xy * xy, 1, True) 161 | dists = xy2 - 2 * tf.matmul(xy, tf.transpose(xy)) + tf.transpose(xy2) 162 | return tf.reduce_mean(dists) 163 | 164 | def output_to_pdist(bx, by): 165 | """ 166 | Tensor op: calculate expectation of box distance given yolo outpput bx & by. 167 | # Args: 168 | bx - YOLOv3 output batch x coordinates in shape (N, 3549, 3, 1) 169 | by - YOLOv3 output batch y coordinates in shape (N, 3549, 3, 1) 170 | """ 171 | bxby = tf.concat([bx, by], axis=-1) 172 | bxby = tf.reshape(bxby, [-1, 2]) 173 | return pdist(bxby) 174 | 175 | def pairwise_IoUs(bs1, bs2): 176 | """ 177 | Tensor op: Calculate pairwise IoUs given two sets of boxes. 178 | # Arguments: 179 | bs1, bs2 - tensor of boxes in shape (?, 4) 180 | # Content: 181 | X11,y11------x12,y11 X21,y21------x22,y21 182 | | | | | 183 | | | | | 184 | x11,y12-------x12,y12 x21,y22-------x22,y22 185 | # Returns: 186 | iou - a tensor of the matrix containing pairwise IoUs, in shape (?, ?) 187 | """ 188 | x11, y11, w1, h1 = tf.split(bs1, 4, axis=1) # (N, 1) 189 | x21, y21, w2, h2 = tf.split(bs2, 4, axis=1) # (N, 1) 190 | x12 = x11 + w1 191 | y12 = y11 + h1 192 | x22 = x21 + w2 193 | y22 = y21 + h2 194 | xA = tf.maximum(x11, tf.transpose(x21)) 195 | yA = tf.maximum(y11, tf.transpose(y21)) 196 | xB = tf.minimum(x12, tf.transpose(x22)) 197 | yB = tf.minimum(y12, tf.transpose(y22)) 198 | # prevent 0 area 199 | interArea = tf.maximum((xB - xA + 1), 0) * tf.maximum((yB - yA + 1), 0) 200 | 201 | boxAArea = (x12 - x11 + 1) * (y12 - y11 + 1) 202 | boxBArea = (x22 - x21 + 1) * (y22 - y21 + 1) 203 | 204 | iou = interArea / (boxAArea + tf.transpose(boxBArea) - interArea) 205 | print('iou', iou) 206 | return iou 207 | 208 | 209 | def expectation_of_IoUs(boxes): 210 | """ 211 | Tensor op: Calculate the expectation given all pairwise IoUs. 212 | # Arguments 213 | boxes - boxes of objects. It takes (?, 4) shaped tensor; 214 | # Returns 215 | expt - expectation of IoUs of box pairs. Scalar tensor. 216 | """ 217 | IoUs = pairwise_IoUs(boxes, boxes) 218 | expt = tf.reduce_mean(IoUs) 219 | return expt 220 | 221 | def expectation_of_IoUs_accross_classes(boxes, box_scores): 222 | """ 223 | Tensor op: Calculate IoU expectation for IoU expectations from different class. 224 | Arguments: 225 | #boxes - (3549, 3, 4) tensor output from yolo net 226 | #box_scores - (N1**2+N2**2+N3**2, 3, 80) tensor 227 | Content: 228 | #box_classes - (N1**2+N2**2+N3**2, 3, 1) tensor 229 | Returns: 230 | #expt_over_all_classes - The IoU expectation of box pairs over all classes. 231 | """ 232 | box_classes = tf.cast(tf.argmax(box_scores, axis=-1), tf.int32, name='box_classes') 233 | class_counts = tf.bincount(box_classes) 234 | dominating_cls = tf.argmax(class_counts) 235 | dominating_cls = tf.cast(dominating_cls, tf.int32) 236 | index = tf.equal(box_classes, dominating_cls) 237 | index = tf.cast(index, tf.int32) 238 | others, dominating_boxes = tf.dynamic_partition(boxes, index, num_partitions=2, name='dynamic_partition') 239 | expt_over_all_classes = expectation_of_IoUs(dominating_boxes) 240 | return expt_over_all_classes 241 | 242 | class Daedalus: 243 | """ 244 | Daedalus adversarial example generator based on the Yolo v3 model. 245 | """ 246 | def __init__(self, sess, model, target_class=ATTACK_CLASS, attack_mode=ATTACK_MODE, img_shape=IMAGE_SHAPE, 247 | batch_size=BATCH_SIZE, confidence=CONFIDENCE, learning_rate=LEARNING_RATE, binary_search_steps=BINARY_SEARCH_STEPS, 248 | max_iterations=MAX_ITERATIONS, abort_early=ABORT_EARLY, initial_consts=INITIAL_consts, boxmin=0, boxmax=1): 249 | 250 | # self.sess = tf_debug.LocalCLIDebugWrapperSession(sess) 251 | self.sess = sess 252 | self.LEARNING_RATE = learning_rate 253 | self.MAX_ITERATIONS = max_iterations 254 | self.BINARY_SEARCH_STEPS = binary_search_steps 255 | self.ABORT_EARLY = abort_early 256 | self.initial_consts = initial_consts 257 | self.batch_size = batch_size 258 | self.repeat = binary_search_steps >= 6 259 | self.yolo_model = model 260 | self.confidence = confidence 261 | self.target_class = target_class 262 | self.attack_mode = attack_mode 263 | 264 | def select_class(target_class, boxes, objectness, box_scores, mode='all'): 265 | box_classes = tf.cast(tf.argmax(box_scores, axis=-1), tf.int32, name='box_classes') 266 | class_counts = tf.bincount(box_classes) 267 | print(class_counts) 268 | if mode == 'all': 269 | selected_boxes = tf.reshape(boxes, [BATCH_SIZE, -1, 4]) 270 | selected_scores = tf.reshape(box_scores, [BATCH_SIZE, -1, CLASS_NUM]) 271 | if objectness == None: 272 | return selected_boxes, None, selected_scores 273 | selected_objectness = tf.reshape(objectness, [BATCH_SIZE, -1, 1]) 274 | return selected_boxes, selected_objectness, selected_scores 275 | elif mode == 'most': 276 | selected_cls = tf.argmax(class_counts) 277 | elif mode == 'least': 278 | class_counts = tf.where(tf.equal(class_counts,0), int(1e6)*tf.ones_like(class_counts, dtype=tf.int32), class_counts) 279 | selected_cls = tf.argmin(class_counts) 280 | elif mode == 'single': 281 | file = 'data/coco_classes.txt' 282 | with open(file) as f: 283 | class_names = f.readlines() 284 | class_names = [c.strip() for c in class_names] 285 | selected_cls = class_names.index(target_class) 286 | selected_cls = tf.cast(selected_cls, tf.int32) 287 | index = tf.equal(box_classes, selected_cls) 288 | index = tf.cast(index, tf.int32) 289 | _, selected_boxes = tf.dynamic_partition(boxes, index, num_partitions=2, name='dynamic_partition') 290 | _, selected_scores = tf.dynamic_partition(box_scores, index, num_partitions=2, name='dynamic_partition') 291 | selected_boxes = tf.reshape(selected_boxes, [BATCH_SIZE, -1, 4]) 292 | selected_scores = tf.reshape(selected_scores, [BATCH_SIZE, -1, CLASS_NUM]) 293 | if objectness == None: 294 | return selected_boxes, None, selected_scores 295 | _, selected_objectness = tf.dynamic_partition(objectness, index, num_partitions=2, name='dynamic_partition') 296 | selected_objectness = tf.reshape(selected_objectness, [BATCH_SIZE, -1, 1]) 297 | return selected_boxes, selected_objectness, selected_scores 298 | 299 | # the perturbation we're going to optimize: 300 | perturbations = tf.Variable(np.zeros((batch_size, 301 | img_shape[0], 302 | img_shape[1], 303 | img_shape[2])), dtype=tf.float32, name='perturbations') 304 | # tf variables to sending data to tf: 305 | self.timgs = tf.Variable(np.zeros((batch_size, 306 | img_shape[0], 307 | img_shape[1], 308 | img_shape[2])), dtype=tf.float32, name='self.timgs') 309 | self.consts = tf.Variable(np.zeros(batch_size), dtype=tf.float32, name='self.consts') 310 | 311 | # and here's what we use to assign them: 312 | self.assign_timgs = tf.placeholder(tf.float32, (batch_size, 313 | img_shape[0], 314 | img_shape[1], 315 | img_shape[2])) 316 | self.assign_consts = tf.placeholder(tf.float32, [batch_size]) 317 | 318 | # Tensor operation: the resulting image, tanh'd to keep bounded from 319 | # boxmin to boxmax: 320 | self.boxmul = (boxmax - boxmin) / 2. 321 | self.boxplus = (boxmin + boxmax) / 2. 322 | self.newimgs = tf.tanh(perturbations + self.timgs) * self.boxmul + self.boxplus 323 | 324 | # Get prediction from the model: 325 | outs = self.yolo_model._yolo(self.newimgs) 326 | # [(N, 13, 13, 3, 85), (N, 26, 26, 3, 85), (N, 52, 52, 3, 85)] 327 | print(outs) 328 | # (N, 3549, 3, 4), (N, 3549, 3, 1), (N, 3549, 3, 80) 329 | boxes, objectness, classprobs = process_output(outs) 330 | boxes, objectness, classprobs = select_class(self.target_class, boxes, objectness, classprobs, mode=self.attack_mode) 331 | print(boxes, objectness, classprobs) 332 | self.bx = boxes[..., 0:1] 333 | self.by = boxes[..., 1:2] 334 | self.bw = boxes[..., 2:3] 335 | self.bh = boxes[..., 3:4] 336 | self.obj_scores = objectness 337 | self.class_probs = classprobs 338 | self.box_scores = tf.multiply(self.obj_scores, tf.reduce_max(self.class_probs, axis=-1, keepdims=True)) 339 | 340 | # Optimisation metrics: 341 | self.l2dist = tf.reduce_sum(tf.square(self.newimgs - (tf.tanh(self.timgs) * self.boxmul + self.boxplus)), [1, 2, 3]) 342 | 343 | # Define DDoS losses: loss must be a tensor here! 344 | # Make the objectness of all detections to be 1. 345 | self.loss1_1_x = tf.reduce_mean(tf.square(self.box_scores - 1), [-2, -1]) 346 | 347 | # Minimising the size of all bounding box. 348 | #self.f1 = tf.reduce_mean(IoU_expts) 349 | self.f3 = tf.reduce_mean(tf.square(tf.multiply(self.bw, self.bh)), [-2, -1]) 350 | #self.f2 = self.f3 + 1/output_to_pdist(self.bx, self.by) 351 | 352 | # add two loss terms together 353 | self.loss_adv = self.loss1_1_x + self.f3 354 | self.loss1 = tf.reduce_mean(self.consts * self.loss_adv) 355 | self.loss2 = tf.reduce_mean(self.l2dist) 356 | self.loss = self.loss1 + self.loss2 357 | 358 | # Setup the adam optimizer and keep track of variables we're creating 359 | start_vars = set(x.name for x in tf.global_variables()) 360 | optimizer = tf.train.AdamOptimizer(self.LEARNING_RATE) 361 | self.train = optimizer.minimize(self.loss, var_list=[perturbations]) 362 | end_vars = tf.global_variables() 363 | new_vars = [x for x in end_vars if x.name not in start_vars] 364 | 365 | # these are the variables to initialize when we run 366 | self.setup = [] 367 | self.setup.append(self.timgs.assign(self.assign_timgs)) 368 | self.setup.append(self.consts.assign(self.assign_consts)) 369 | self.init = tf.variables_initializer(var_list=[perturbations] + new_vars) 370 | 371 | def attack_batch(self, imgs): 372 | """ 373 | Run the attack on a batch of images and labels. 374 | """ 375 | 376 | def check_success(loss, init_loss): 377 | """ 378 | Check if the initial loss value has been reduced by 'self.confidence' percent 379 | """ 380 | return loss <= init_loss * (1 - self.confidence) 381 | 382 | batch_size = self.batch_size 383 | 384 | # convert images to arctanh-space 385 | imgs = np.arctanh((imgs - self.boxplus) / self.boxmul * 0.999999) 386 | 387 | # set the lower and upper bounds of the constsant. 388 | lower_bound = np.zeros(batch_size) 389 | consts = np.ones(batch_size) * self.initial_consts 390 | upper_bound = np.ones(batch_size) * 1e10 391 | 392 | # store the best l2, score, and image attack 393 | o_bestl2 = [1e10] * batch_size 394 | o_bestloss = [1e10] * batch_size 395 | o_bestattack = [np.zeros(imgs[0].shape)] * batch_size 396 | 397 | for outer_step in range(self.BINARY_SEARCH_STEPS): 398 | # completely reset adam's internal state. 399 | self.sess.run(self.init) 400 | 401 | # take in the current data batch. 402 | batch = imgs[:batch_size] 403 | 404 | # cache the current best l2 and score. 405 | bestl2 = [1e10] * batch_size 406 | # bestconfidence = [-1]*batch_size 407 | bestloss = [1e10] * batch_size 408 | 409 | # The last iteration (if we run many steps) repeat the search once. 410 | if self.repeat and outer_step == self.BINARY_SEARCH_STEPS - 1: 411 | consts = upper_bound 412 | 413 | # set the variables so that we don't have to send them over again. 414 | self.sess.run(self.setup, {self.assign_timgs: batch, 415 | self.assign_consts: consts}) 416 | 417 | # start gradient descent attack 418 | print('adjust c to:', sess.run(self.consts)) 419 | init_loss = sess.run(self.loss) 420 | init_adv_losses = sess.run(self.loss_adv) 421 | prev = init_loss * 1.1 422 | for iteration in range(self.MAX_ITERATIONS): 423 | # perform the attack on a single example 424 | _, l, l2s, l1s, nimgs, c = self.sess.run([self.train, self.loss, self.l2dist, self.loss_adv, self.newimgs, self.consts]) 425 | # print out the losses every 10% 426 | if iteration % (self.MAX_ITERATIONS // 10) == 0: 427 | print('===iteration:', iteration, '===') 428 | print('attacked box number:', sess.run(self.bw).shape) 429 | print('loss values of box confidence and dimension:', sess.run([self.loss1_1_x, self.f3])) 430 | print('adversarial losses:', l1s) 431 | print('distortions:', l2s) 432 | 433 | # check if we should abort search if we're getting nowhere. 434 | if self.ABORT_EARLY and iteration % (self.MAX_ITERATIONS // 10) == 0: 435 | if l > prev * .9999: 436 | break 437 | prev = l 438 | 439 | # update the best result found so far 440 | for e, (l1, l2, ii) in enumerate(zip(l1s, l2s, nimgs)): 441 | if l2 < bestl2[e] and check_success(l1, init_adv_losses[e]): 442 | bestl2[e] = l2 443 | bestloss[e] = l1 444 | if l2 < o_bestl2[e] and check_success(l1, init_adv_losses[e]): 445 | o_bestl2[e] = l2 446 | o_bestloss[e] = l1 447 | o_bestattack[e] = ii 448 | 449 | # adjust the constsant as needed 450 | for e in range(batch_size): 451 | if check_success(l1s[e], init_adv_losses[e]): 452 | # success, divide consts by two 453 | upper_bound[e] = min(upper_bound[e], consts[e]) 454 | if upper_bound[e] < 1e9: 455 | consts[e] = (lower_bound[e] + upper_bound[e]) / 2 456 | else: 457 | # failure, either multiply by 10 if no solution found yet 458 | # or do binary search with the known upper bound 459 | lower_bound[e] = max(lower_bound[e], consts[e]) 460 | if upper_bound[e] < 1e9: 461 | consts[e] = (lower_bound[e] + upper_bound[e]) / 2 462 | else: 463 | consts[e] *= 10 464 | # return the best solution found 465 | o_bestl2 = np.array(o_bestl2) 466 | return o_bestattack, o_bestl2 467 | 468 | 469 | def attack(self, imgs): 470 | """ 471 | Perform the L_2 attack on the given images for the given targets. 472 | If self.targeted is true, then the targets represents the target labels. 473 | If self.targeted is false, then targets are the original class labels. 474 | """ 475 | r = [] 476 | ds = [] 477 | print('go up to', len(imgs)) 478 | for i in range(0, len(imgs), self.batch_size): 479 | print('tick', i) 480 | X_adv, dists = self.attack_batch(imgs[i:i + self.batch_size]) 481 | path = SAVE_PATH+'{0} confidence'.format(self.confidence) 482 | if not os.path.exists(path): 483 | os.makedirs(path) 484 | np.save(path+'/Distortions of images {0} to {1}.npy'.format(i, i+self.batch_size), dists) 485 | for j in range(len(X_adv)): 486 | io.imsave(path+'/Best example of {1} Distortion {2}.png'.format(self.confidence, i+j, dists[j]), X_adv[j]) 487 | r.extend(X_adv) 488 | ds.extend(dists) 489 | return np.array(r), np.array(ds) 490 | 491 | 492 | if __name__ == '__main__': 493 | sess = tf.InteractiveSession() 494 | init = tf.global_variables_initializer() 495 | sess.run(init) 496 | ORACLE = YOLO(0.6, 0.5) # The auguments do not matter. 497 | X_test = [] 498 | for (root, dirs, files) in os.walk('../Datasets/COCO/val2017/'): 499 | if files: 500 | for f in files: 501 | print(f) 502 | path = os.path.join(root, f) 503 | image = cv2.imread(path) 504 | image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB) # RGB 505 | image = process_image(image) 506 | #image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR) 507 | X_test.append(image) 508 | EXAMPLE_NUM -= 1 509 | if EXAMPLE_NUM == 0: 510 | break 511 | X_test = np.concatenate(X_test, axis=0) 512 | attacker = Daedalus(sess, ORACLE) 513 | X_adv, distortions = attacker.attack(X_test) 514 | np.savez(SAVE_PATH+'{0} confidence/Daedalus example batch.npz'.format(CONFIDENCE), X_adv=X_adv, distortions=distortions) 515 | writer = tf.summary.FileWriter("log", sess.graph) 516 | writer.close() -------------------------------------------------------------------------------- /model/darknet53.py: -------------------------------------------------------------------------------- 1 | """Darknet-53 for yolo v3. 2 | """ 3 | from keras.models import Model 4 | from keras.layers import Input, Conv2D, GlobalAveragePooling2D, Dense 5 | from keras.layers import add, Activation, BatchNormalization 6 | from keras.layers.advanced_activations import LeakyReLU 7 | from keras.regularizers import l2 8 | 9 | 10 | def conv2d_unit(x, filters, kernels, strides=1): 11 | """Convolution Unit 12 | This function defines a 2D convolution operation with BN and LeakyReLU. 13 | 14 | # Arguments 15 | x: Tensor, input tensor of conv layer. 16 | filters: Integer, the dimensionality of the output space. 17 | kernels: An integer or tuple/list of 2 integers, specifying the 18 | width and height of the 2D convolution window. 19 | strides: An integer or tuple/list of 2 integers, 20 | specifying the strides of the convolution along the width and 21 | height. Can be a single integer to specify the same value for 22 | all spatial dimensions. 23 | 24 | # Returns 25 | Output tensor. 26 | """ 27 | x = Conv2D(filters, kernels, 28 | padding='same', 29 | strides=strides, 30 | activation='linear', 31 | kernel_regularizer=l2(5e-4))(x) 32 | x = BatchNormalization()(x) 33 | x = LeakyReLU(alpha=0.1)(x) 34 | 35 | return x 36 | 37 | 38 | def residual_block(inputs, filters): 39 | """Residual Block 40 | This function defines a 2D convolution operation with BN and LeakyReLU. 41 | 42 | # Arguments 43 | x: Tensor, input tensor of residual block. 44 | kernels: An integer or tuple/list of 2 integers, specifying the 45 | width and height of the 2D convolution window. 46 | 47 | # Returns 48 | Output tensor. 49 | """ 50 | x = conv2d_unit(inputs, filters, (1, 1)) 51 | x = conv2d_unit(x, 2 * filters, (3, 3)) 52 | x = add([inputs, x]) 53 | x = Activation('linear')(x) 54 | 55 | return x 56 | 57 | 58 | def stack_residual_block(inputs, filters, n): 59 | """Stacked residual Block 60 | """ 61 | x = residual_block(inputs, filters) 62 | 63 | for i in range(n - 1): 64 | x = residual_block(x, filters) 65 | 66 | return x 67 | 68 | 69 | def darknet_base(inputs): 70 | """Darknet-53 base model. 71 | """ 72 | 73 | x = conv2d_unit(inputs, 32, (3, 3)) 74 | 75 | x = conv2d_unit(x, 64, (3, 3), strides=2) 76 | x = stack_residual_block(x, 32, n=1) 77 | 78 | x = conv2d_unit(x, 128, (3, 3), strides=2) 79 | x = stack_residual_block(x, 64, n=2) 80 | 81 | x = conv2d_unit(x, 256, (3, 3), strides=2) 82 | x = stack_residual_block(x, 128, n=8) 83 | 84 | x = conv2d_unit(x, 512, (3, 3), strides=2) 85 | x = stack_residual_block(x, 256, n=8) 86 | 87 | x = conv2d_unit(x, 1024, (3, 3), strides=2) 88 | x = stack_residual_block(x, 512, n=4) 89 | 90 | return x 91 | 92 | 93 | def darknet(): 94 | """Darknet-53 classifier. 95 | """ 96 | inputs = Input(shape=(416, 416, 3)) 97 | x = darknet_base(inputs) 98 | 99 | x = GlobalAveragePooling2D()(x) 100 | x = Dense(1000, activation='softmax')(x) 101 | 102 | model = Model(inputs, x) 103 | 104 | return model 105 | 106 | 107 | if __name__ == '__main__': 108 | model = darknet() 109 | print(model.summary()) 110 | -------------------------------------------------------------------------------- /model/yolo_model.py: -------------------------------------------------------------------------------- 1 | """YOLO v3 output 2 | """ 3 | import numpy as np 4 | import keras.backend as K 5 | from keras.models import load_model 6 | 7 | class YOLO: 8 | def __init__(self, obj_threshold, nms_threshold): 9 | """Init. 10 | 11 | # Arguments 12 | obj_threshold: Integer, threshold for object. 13 | nms_threshold: Integer, threshold for box. 14 | """ 15 | self._t1 = obj_threshold 16 | self._t2 = nms_threshold 17 | self._yolo = load_model('../YOLOv3/data/yolo.h5') 18 | 19 | def _process_feats(self, out, anchors, mask): 20 | """process output features. 21 | 22 | # Arguments 23 | out: Tensor (N, N, 3, 4 + 1 +80), output feature map of yolo. 24 | anchors: List, anchors for box. 25 | mask: List, mask for anchors. 26 | 27 | # Returns 28 | boxes: ndarray (N, N, 3, 4), x,y,w,h for per box. 29 | box_confidence: ndarray (N, N, 3, 1), confidence for per box. 30 | box_class_probs: ndarray (N, N, 3, 80), class probs for per box. 31 | """ 32 | grid_h, grid_w, num_boxes = map(int, out.shape[1: 4]) 33 | 34 | anchors = [anchors[i] for i in mask] 35 | # Reshape to batch, height, width, num_anchors, box_params. 36 | anchors_tensor = K.reshape(K.variable(anchors), 37 | [1, 1, len(anchors), 2]) 38 | out = out[0] 39 | box_xy = K.get_value(K.sigmoid(out[..., :2])) 40 | box_wh = K.get_value(K.exp(out[..., 2:4]) * anchors_tensor) 41 | box_confidence = K.get_value(K.sigmoid(out[..., 4])) 42 | box_confidence = np.expand_dims(box_confidence, axis=-1) 43 | box_class_probs = K.get_value(K.sigmoid(out[..., 5:])) 44 | 45 | col = np.tile(np.arange(0, grid_w), grid_w).reshape(-1, grid_w) 46 | row = np.tile(np.arange(0, grid_h).reshape(-1, 1), grid_h) 47 | 48 | col = col.reshape(grid_h, grid_w, 1, 1).repeat(3, axis=-2) 49 | row = row.reshape(grid_h, grid_w, 1, 1).repeat(3, axis=-2) 50 | grid = np.concatenate((col, row), axis=-1) 51 | 52 | box_xy += grid 53 | box_xy /= (grid_w, grid_h) 54 | box_wh /= (416, 416) 55 | box_xy -= (box_wh / 2.) 56 | boxes = np.concatenate((box_xy, box_wh), axis=-1) 57 | #print('box confidences are: {0}',format(box_confidence)) 58 | return boxes, box_confidence, box_class_probs 59 | 60 | def _filter_boxes(self, boxes, box_confidences, box_class_probs): 61 | """Filter boxes with object threshold. 62 | 63 | # Arguments 64 | boxes: ndarray, boxes of objects. 65 | box_confidences: ndarray, confidences of objects. 66 | box_class_probs: ndarray, class_probs of objects. 67 | 68 | # Returns 69 | boxes: ndarray, filtered boxes. 70 | classes: ndarray, classes for boxes. 71 | scores: ndarray, scores for boxes. 72 | """ 73 | box_scores = box_confidences * box_class_probs 74 | box_classes = np.argmax(box_scores, axis=-1) 75 | box_class_scores = np.max(box_scores, axis=-1) 76 | pos = np.where(box_class_scores >= self._t1) 77 | 78 | boxes = boxes[pos] 79 | classes = box_classes[pos] 80 | scores = box_class_scores[pos] 81 | 82 | return boxes, classes, scores 83 | 84 | def _soft_nms(self, boxes, scores): 85 | """Suppress non-maximal boxes. 86 | 87 | # Arguments 88 | boxes: ndarray, boxes of objects. 89 | scores: ndarray, scores of objects. 90 | 91 | # Returns 92 | keep: ndarray, index of effective boxes. 93 | """ 94 | x = boxes[:, 0] 95 | y = boxes[:, 1] 96 | w = boxes[:, 2] 97 | h = boxes[:, 3] 98 | 99 | areas = w * h 100 | order = scores.argsort()[::-1] 101 | keep = [] 102 | while order.size > 0: 103 | i = order[0] 104 | keep.append(i) 105 | order = order[1:] 106 | 107 | xx1 = np.maximum(x[i], x[order]) 108 | yy1 = np.maximum(y[i], y[order]) 109 | xx2 = np.minimum(x[i] + w[i], x[order] + w[order]) 110 | yy2 = np.minimum(y[i] + h[i], y[order] + h[order]) 111 | 112 | w1 = np.maximum(0.0, xx2 - xx1 + 1) 113 | h1 = np.maximum(0.0, yy2 - yy1 + 1) 114 | inter = w1 * h1 115 | 116 | ovr = inter / (areas[i] + areas[order] - inter) 117 | 118 | scores[order] *= np.exp((ovr**2)/(-0.5)) # change scores based on IoU 119 | scores = scores[order] 120 | order = scores.argsort()[::-1] # re-order 121 | 122 | keep = np.array(keep) 123 | return keep 124 | 125 | def _nms_boxes(self, boxes, scores): 126 | """Suppress non-maximal boxes. 127 | 128 | # Arguments 129 | boxes: ndarray, boxes of objects. 130 | scores: ndarray, scores of objects. 131 | 132 | # Returns 133 | keep: ndarray, index of effective boxes. 134 | """ 135 | x = boxes[:, 0] 136 | y = boxes[:, 1] 137 | w = boxes[:, 2] 138 | h = boxes[:, 3] 139 | 140 | areas = w * h 141 | order = scores.argsort()[::-1] 142 | 143 | keep = [] 144 | while order.size > 0: 145 | i = order[0] 146 | keep.append(i) 147 | 148 | xx1 = np.maximum(x[i], x[order[1:]]) 149 | yy1 = np.maximum(y[i], y[order[1:]]) 150 | xx2 = np.minimum(x[i] + w[i], x[order[1:]] + w[order[1:]]) 151 | yy2 = np.minimum(y[i] + h[i], y[order[1:]] + h[order[1:]]) 152 | 153 | w1 = np.maximum(0.0, xx2 - xx1 + 1) 154 | h1 = np.maximum(0.0, yy2 - yy1 + 1) 155 | inter = w1 * h1 156 | 157 | ovr = inter / (areas[i] + areas[order[1:]] - inter) 158 | inds = np.where(ovr <= self._t2)[0] 159 | order = order[inds + 1] 160 | 161 | keep = np.array(keep) 162 | 163 | return keep 164 | 165 | def _yolo_out(self, outs, shape): 166 | """Process output of yolo base net. 167 | 168 | # Argument: 169 | outs: output of yolo base net. 170 | shape: shape of original image. 171 | 172 | # Returns: 173 | boxes: ndarray, boxes of objects. 174 | classes: ndarray, classes of objects. 175 | scores: ndarray, scores of objects. 176 | """ 177 | masks = [[6, 7, 8], [3, 4, 5], [0, 1, 2]] 178 | anchors = [[10, 13], [16, 30], [33, 23], [30, 61], [62, 45], 179 | [59, 119], [116, 90], [156, 198], [373, 326]] 180 | 181 | boxes, classes, scores = [], [], [] 182 | 183 | for out, mask in zip(outs, masks): 184 | b, c, s = self._process_feats(out, anchors, mask) 185 | b, c, s = self._filter_boxes(b, c, s) 186 | boxes.append(b) 187 | classes.append(c) 188 | scores.append(s) 189 | 190 | boxes = np.concatenate(boxes) 191 | classes = np.concatenate(classes) 192 | scores = np.concatenate(scores) 193 | 194 | 195 | 196 | # Scale boxes back to original image shape. 197 | width, height = shape[1], shape[0] 198 | image_dims = [width, height, width, height] 199 | boxes = boxes * image_dims 200 | 201 | nboxes, nclasses, nscores = [], [], [] 202 | for c in set(classes): 203 | inds = np.where(classes == c) 204 | b = boxes[inds] 205 | c = classes[inds] 206 | s = scores[inds] 207 | 208 | keep = self._nms_boxes(b, s) #use NMS 209 | #keep = self._soft_nms(b, s) #use soft-NMS 210 | 211 | nboxes.append(b[keep]) 212 | nclasses.append(c[keep]) 213 | nscores.append(s[keep]) 214 | 215 | if not nclasses and not nscores: 216 | return None, None, None 217 | 218 | boxes = np.concatenate(nboxes) 219 | classes = np.concatenate(nclasses) 220 | scores = np.concatenate(nscores) 221 | 222 | return boxes, classes, scores 223 | 224 | def predict(self, image, shape): 225 | """Detect the objects with yolo. 226 | 227 | # Arguments 228 | image: ndarray, processed input image. 229 | shape: shape of original image. 230 | 231 | # Returns 232 | boxes: ndarray, boxes of objects. 233 | classes: ndarray, classes of objects. 234 | scores: ndarray, scores of objects. 235 | """ 236 | 237 | raw_outs = self._yolo.predict(image) 238 | boxes, classes, scores = self._yolo_out(raw_outs, shape) 239 | 240 | return boxes, classes, scores, raw_outs 241 | -------------------------------------------------------------------------------- /resources/l2attack.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/NeuralSec/Daedalus-attack/3f9cb38f6389cb7cd4895a2f9679aa5ce3c81e70/resources/l2attack.jpg --------------------------------------------------------------------------------