├── README.md ├── data_prepare.py ├── evaluate&predict.py └── train.py /README.md: -------------------------------------------------------------------------------- 1 | # Prediction-Based-GNSS-Spoofing-Attack-Detection-for-Autonomous-Vehicle 2 | Experimental realization for Prediction-Based GNSS Spoofing Attack Detection for Autonomous Vehicle in python 3 | more infomation see https://blog.csdn.net/weixin_44546393/article/details/114264497 4 | -------------------------------------------------------------------------------- /data_prepare.py: -------------------------------------------------------------------------------- 1 | import os 2 | import numpy as np 3 | import pandas as pd 4 | import pyproj 5 | from scipy.interpolate import make_interp_spline 6 | import math 7 | import matplotlib.pyplot as plt 8 | # a = 6378137 9 | # b = 6356752.3142 10 | # esq = 6.69437999014 * 0.001 11 | # e1sq = 6.73949674228 * 0.001 12 | # def ecef2geodetic(ecef, radians=False): 13 | # """ 14 | # Convert ECEF coordinates to geodetic using ferrari's method 15 | # """ 16 | # # Save shape and export column 17 | # ecef = np.atleast_1d(ecef) 18 | # input_shape = ecef.shape 19 | # ecef = np.atleast_2d(ecef) 20 | # x, y, z = ecef[:, 0], ecef[:, 1], ecef[:, 2] 21 | # 22 | # ratio = 1.0 if radians else (180.0 / np.pi) 23 | # 24 | # # Conver from ECEF to geodetic using Ferrari's methods 25 | # # https://en.wikipedia.org/wiki/Geographic_coordinate_conversion#Ferrari.27s_solution 26 | # r = np.sqrt(x * x + y * y) 27 | # Esq = a * a - b * b 28 | # F = 54 * b * b * z * z 29 | # G = r * r + (1 - esq) * z * z - esq * Esq 30 | # C = (esq * esq * F * r * r) / (pow(G, 3)) 31 | # S = np.cbrt(1 + C + np.sqrt(C * C + 2 * C)) 32 | # P = F / (3 * pow((S + 1 / S + 1), 2) * G * G) 33 | # Q = np.sqrt(1 + 2 * esq * esq * P) 34 | # r_0 = -(P * esq * r) / (1 + Q) + np.sqrt(0.5 * a * a*(1 + 1.0 / Q) - \ 35 | # P * (1 - esq) * z * z / (Q * (1 + Q)) - 0.5 * P * r * r) 36 | # U = np.sqrt(pow((r - esq * r_0), 2) + z * z) 37 | # V = np.sqrt(pow((r - esq * r_0), 2) + (1 - esq) * z * z) 38 | # Z_0 = b * b * z / (a * V) 39 | # h = U * (1 - b * b / (a * V)) 40 | # lat = ratio*np.arctan((z + e1sq * Z_0) / r) 41 | # lon = ratio*np.arctan2(y, x) 42 | # 43 | # # stack the new columns and return to the original shape 44 | # geodetic = np.column_stack((lat, lon, h)) 45 | # return geodetic.reshape(input_shape) 46 | # 在字符串指定位置的插入字符 47 | def str_insert(str_origin, pos, str_add): 48 | str_list = list(str_origin) # 字符串转list 49 | str_list.insert(pos, str_add) # 在指定位置插入字符串 50 | str_out = ''.join(str_list) # 空字符连接 51 | return str_out 52 | 53 | # 数组去重 54 | def unique(old_list): 55 | newList = [] 56 | # 判断相邻时间是否相等 57 | if np.any(old_list[1:] == old_list[:-1]): 58 | for x in old_list: 59 | if x in newList: 60 | # 若相等，则加上一个微小的数使其不等 61 | x = x + 0.005 62 | newList.append(x) 63 | return np.array(newList) 64 | else: return old_list 65 | 66 | def series_to_supervised(data, n_in=1, n_out=1, dropnan=True): 67 | n_vars = 1 if type(data) is list else data.shape[1] 68 | df = pd.DataFrame(data) 69 | column_names = ['lats', 'lons', 'CAN_speeds', 'steering_angles', 'acceleration_forward'] 70 | cols, names = list(), list() 71 | # input sequence (t-n, ... t-1) 72 | for i in range(n_in, 0, -1): 73 | cols.append(df.shift(i)) 74 | names += [('%s(t-%d)' % (j, i)) for j in column_names] 75 | # forecast sequence (t, t+1, ... t+n) 76 | for i in range(0, n_out): 77 | cols.append(df.shift(-i)) 78 | if i == 0: 79 | names += [('%s(t)' % (j)) for j in column_names] 80 | else: 81 | names += [('%s(t+%d)' % (j, i)) for j in column_names] 82 | # put it all together 83 | agg = pd.concat(cols, axis=1) 84 | agg.columns = names 85 | # drop rows with NaN values 86 | if dropnan: 87 | agg.dropna(inplace=True) 88 | return agg 89 | 90 | EARTH_REDIUS = 6378.137 91 | 92 | def rad(d): 93 | return d * math.pi / 180.0 94 | def getDistance(lat1, lng1, lat2, lng2): 95 | # 对数组取元素做运算 96 | res = [] 97 | for i in range(len(lat1)): 98 | radLat1 = rad(lat1[i]) 99 | radLat2 = rad(lat2[i]) 100 | a = radLat1 - radLat2 101 | b = rad(lng1[i]) - rad(lng2[i]) 102 | s = 2 * math.asin(math.sqrt(math.pow(math.sin(a / 2), 2) + math.cos(radLat1) * math.cos(radLat2) * math.pow( 103 | math.sin(b / 2), 2))) 104 | s = s * EARTH_REDIUS * 1000 105 | res.append(s) 106 | return res 107 | 108 | if __name__ == '__main__': 109 | 110 | # 用于GNSS坐标转化 111 | position_transformer = pyproj.Transformer.from_crs( 112 | {"proj": 'geocent', "ellps": 'WGS84', "datum": 'WGS84'}, 113 | {"proj": 'latlong', "ellps": 'WGS84', "datum": 'WGS84'}, 114 | ) 115 | dataset_directory = 'D:\comma2k19' 116 | chunk_set = [] 117 | for chunk in os.listdir(dataset_directory): 118 | # 忽略生成的csv文件 119 | if ".csv" in chunk: 120 | continue 121 | # 如果序号为单个时在前补零，以便后面排序 122 | if len(chunk) == 7: 123 | used_name = chunk 124 | chunk = str_insert(chunk,6,'0') 125 | os.rename(os.path.join(dataset_directory, used_name), os.path.join(dataset_directory, chunk)) 126 | chunk_set.append(os.path.join(dataset_directory, chunk)) 127 | # 将序号小的片段放在前面 128 | chunk_set.sort() 129 | # 选一个chunk来训练（200分钟） 130 | chunk_index = 0 131 | route_set = [] 132 | for route_id in os.listdir(chunk_set[chunk_index]): 133 | # 忽略生成的csv文件 134 | if ".csv" in route_id: 135 | continue 136 | route_set.append(os.path.join(chunk_set[chunk_index], route_id)) 137 | segment_set = [] 138 | # 选一个路段训练 139 | route_index = 9 140 | for segment in os.listdir(route_set[route_index]): 141 | # 如果序号为单个时在前补零，以便后面排序 142 | if len(segment) == 1: 143 | used_name = segment 144 | segment = '0'+segment 145 | os.rename(os.path.join(route_set[route_index], used_name),os.path.join(route_set[route_index], segment)) 146 | segment_set.append(os.path.join(route_set[route_index], segment)) 147 | # 将序号小的片段放在前面 148 | segment_set.sort() 149 | times = [] 150 | lons = [] 151 | lats = [] 152 | orientations = [] 153 | CAN_speeds = [] 154 | steering_angles = [] 155 | acceleration_forward = [] 156 | for main_dir in segment_set: 157 | # 导入GNSS的时间和位置(pose)并将位置转化为经纬度 158 | temp_GNSS_time = np.load(main_dir + '\\global_pose\\frame_times') 159 | times = np.append(times, temp_GNSS_time) 160 | # 打印每一段的长度 161 | print(len(temp_GNSS_time)) 162 | positions = np.load(main_dir + '\\global_pose\\frame_positions') 163 | positions = position_transformer.transform(positions[:, 0], positions[:, 1], positions[:, 2], radians=False) 164 | lats = np.append(lats, positions[1]) 165 | lons = np.append(lons, positions[0]) 166 | # Conver from ECEF to geodetic using Ferrari's methods 167 | # positions = ecef2geodetic(positions) 168 | # lats = np.append(lats, positions[:, 0]) 169 | # lons = np.append(lons, positions[:, 1]) 170 | # 暂时不用orientation 171 | # orientation = np.load(main_dir + '\\global_pose\\frame_orientations') 172 | # orientations = np.append(orientations, np.load(main_dir + '\\global_pose\\frame_orientations')) 173 | temp_CAN_times = np.load(main_dir + '\\processed_log\\CAN\\speed\\t') 174 | # 确保时间无重复值 175 | temp_CAN_speed_times = unique(temp_CAN_times) 176 | # 对CAN数据按照GNSS参考时间插值 177 | temp_CAN_speeds = make_interp_spline(temp_CAN_speed_times, np.load(main_dir + '\\processed_log\\CAN\\speed\\value'))(temp_GNSS_time).flatten() 178 | CAN_speeds = np.append(CAN_speeds, temp_CAN_speeds) 179 | # CAN_angles_times和CAN_speed_times有时不一致 180 | temp_CAN_angles_times = np.load(main_dir + '\\processed_log\\CAN\\steering_angle\\t') 181 | temp_steering_angles = np.load(main_dir + '\\processed_log\\CAN\\steering_angle\\value') 182 | temp_CAN_angles_times = unique(temp_CAN_angles_times) 183 | temp_steering_angles = make_interp_spline(temp_CAN_angles_times, temp_steering_angles)(temp_GNSS_time) 184 | steering_angles = np.append(steering_angles, temp_steering_angles) 185 | # 对IMU数据按照GNSS参考时间插值 186 | temp_IMU_times = np.load(main_dir + '\\processed_log\\IMU\\accelerometer\\t') 187 | temp_acceleration_forward = make_interp_spline(temp_IMU_times, np.load(main_dir + 188 | '\\processed_log\\IMU\\accelerometer\\value')[:, 0])(temp_GNSS_time) 189 | acceleration_forward = np.append(acceleration_forward, temp_acceleration_forward) 190 | 191 | DataSet = list(zip(times, lats, lons, CAN_speeds, steering_angles, acceleration_forward)) 192 | column_names = ['times', 'lats', 'lons', 'CAN_speeds', 'steering_angles', 'acceleration_forward'] 193 | df = pd.DataFrame(data=DataSet, columns=column_names) 194 | times = df['times'].values 195 | df = df.set_index(['times'], drop=True) 196 | values = df.values.astype('float64') 197 | # 转为监督学习问题 198 | reframed = series_to_supervised(values, 1, 1) 199 | # 计算距离 200 | lons_t = reframed['lons(t)'].values 201 | lats_t = reframed['lats(t)'].values 202 | distance = np.array(getDistance(lats[:-1], lons[:-1], lats_t, lons_t)) 203 | # drop columns we don't want to predict including（CAN_speed,steering_angel, acceleration_forward) 204 | reframed.drop(reframed.columns[[0, 1, 5, 6, 7, 8, 9]], axis=1, inplace=True) 205 | # 时间和计算的距离添加到数据集 206 | reframed['distance'] = distance 207 | reframed['times'] = times[: -1] 208 | # for i in distance: 209 | # if i > 100: 210 | # print(i) 211 | plt.plot(times[:-1], distance) 212 | plt.xlabel('Boot time (s)', fontsize=18) 213 | plt.ylabel('Distance travelled during single timestamp (m) ', fontsize=12) 214 | plt.show() 215 | # 将合并的数据集保存到.csv文件中 216 | reframed.to_csv(route_set[route_index]+".csv", index=False, sep=',') 217 | 218 | 219 | -------------------------------------------------------------------------------- /evaluate&predict.py: -------------------------------------------------------------------------------- 1 | import pandas as pd 2 | import math 3 | from sklearn.metrics import mean_squared_error 4 | import matplotlib.pyplot as plt 5 | import tensorflow as tf 6 | from sklearn.preprocessing import MinMaxScaler 7 | gpus = tf.config.experimental.list_physical_devices('GPU') 8 | for gpu in gpus: 9 | tf.config.experimental.set_memory_growth(gpu, True) 10 | def average(seq, total=0.0): 11 | num = 0 12 | for item in seq: 13 | total += item 14 | num += 1 15 | return total / num 16 | 17 | if __name__ == '__main__': 18 | 19 | CSV_FILE_PATH = 'D:\\comma2k19\\Chunk_03\\99c94dc769b5d96e_2018-05-01--08-13-53.csv' 20 | df = pd.read_csv(CSV_FILE_PATH) 21 | values = df.to_numpy() 22 | times = values[:, -1] 23 | distance = values[:, -2] 24 | model = tf.keras.models.load_model('lstm.model') 25 | test_X = values[:, :3] 26 | # 因为训练的时候输入特征是归一化的，所以预测的时候也要将输入特征归一化 27 | scaler = MinMaxScaler(feature_range=(0, 1)) 28 | test_X = scaler.fit_transform(test_X) 29 | test_X = test_X.reshape((test_X.shape[0], 1, test_X.shape[1])) 30 | # train_len = (int)(0.75 * len(values[:, 0])) 31 | # train = values[:train_len, :] 32 | # test = values[train_len:, :] 33 | test_y = distance 34 | yhat = model.predict(test_X)[:, 0] 35 | rmse = math.sqrt(mean_squared_error(yhat, test_y)) 36 | print('Test RMSE: %.3f' % rmse) 37 | scores = model.evaluate(test_X, test_y) 38 | rmse = math.sqrt(mean_squared_error(yhat, test_y)) 39 | plt.plot(times, yhat, label='prediction') 40 | plt.plot(times, distance, label="ground_truth") 41 | plt.title('Comparison between truth and prediction', fontsize=18) 42 | plt.xlabel('Boot time (s)', fontsize=18) 43 | plt.ylabel('Distance travelled during single timestamp (m) ', fontsize=12) 44 | plt.legend() 45 | plt.show() 46 | min = min((distance - yhat), key=abs) 47 | max = max((distance - yhat), key=abs) 48 | avr = average(distance-yhat) 49 | print('Min:%f' % min) 50 | print('Max:%f' % max) 51 | print('average:%f' % avr) 52 | -------------------------------------------------------------------------------- /train.py: -------------------------------------------------------------------------------- 1 | import pandas as pd 2 | import math 3 | import numpy as np 4 | from sklearn.preprocessing import MinMaxScaler 5 | from keras.models import Sequential 6 | from keras.layers import Dense 7 | from keras.layers import LSTM 8 | from sklearn.metrics import mean_squared_error 9 | import matplotlib.pyplot as plt 10 | from keras import optimizers 11 | import tensorflow as tf 12 | gpus = tf.config.experimental.list_physical_devices('GPU') 13 | for gpu in gpus: 14 | tf.config.experimental.set_memory_growth(gpu, True) 15 | 16 | 17 | 18 | if __name__ == '__main__': 19 | 20 | train_CSV_FILE_PATH = 'D:\\comma2k19\\Chunk_01\\b0c9d2329ad1606b_2018-08-02--08-34-47.csv' 21 | test_CSV_FILE_PATH = 'D:\\comma2k19\\Chunk_01\\b0c9d2329ad1606b_2018-08-01--21-13-49.csv' 22 | train_df = pd.read_csv(train_CSV_FILE_PATH) 23 | test_df = pd.read_csv(test_CSV_FILE_PATH) 24 | train_values = train_df.to_numpy() 25 | train_times = train_values[:, -1] 26 | train_distance = train_values[:, -2] 27 | test_values = test_df.to_numpy() 28 | test_times = test_values[:, -1] 29 | test_distance = test_values[:, -2] 30 | # 将输入特征归一化 31 | scaler = MinMaxScaler(feature_range=(0, 1)) 32 | train_X, train_y = scaler.fit_transform(train_values[:, :-2]), train_distance 33 | test_X, test_y = scaler.fit_transform(test_values[:, :-2]), test_distance 34 | # # 将四分之三作为训练集 35 | # train_len = len(times) 36 | # train = values[:train_len, :] 37 | # test = values[train_len:, :] 38 | # 划分输入（CAN_speed,steering_angel, acceleration_forward）输出（distance) 39 | # train_X, train_y = train, distance[:train_len] 40 | # test_X, test_y = test, distance[train_len:] 41 | # 将输入（X）改造为LSTM的输入格式，即[samples, timesteps, features] 42 | train_X = train_X.reshape((train_X.shape[0], 1, train_X.shape[1])) 43 | test_X = test_X.reshape((test_X.shape[0], 1, test_X.shape[1])) 44 | 45 | # 设计网络 46 | model = Sequential() 47 | model.add(LSTM(50, input_shape=(train_X.shape[1], train_X.shape[2]))) 48 | model.add(Dense(1)) 49 | # 设置学习率等参数 50 | # adam = optimizers.Adam(lr=0.01, beta_1=0.9, beta_2=0.999, epsilon=1e-08) 51 | model.compile(loss='mae', optimizer='adam') 52 | # fit network 53 | history = model.fit(train_X, train_y, epochs=100, batch_size=50, validation_data=(test_X, test_y), verbose=2, 54 | shuffle=False) 55 | model.save('lstm.model') 56 | # full_X = values[:, :3] 57 | # full_X = full_X.reshape((full_X.shape[0], 1, full_X.shape[1])) 58 | train_yhat = model.predict(train_X)[:, 0] 59 | test_yhat = model.predict(test_X)[:, 0] 60 | rmse = math.sqrt(mean_squared_error(test_yhat, test_y)) 61 | print('Test RMSE: %.3f' % rmse) 62 | # plot history 63 | plt.plot(history.history['loss'], label='train') 64 | plt.plot(history.history['val_loss'], label='test') 65 | # plt.plot(times, yhat, label='prediction') 66 | # plt.plot(times, distance, label="ground_truth") 67 | # plt.title('Comparison between truth and prediction', fontsize=18) 68 | # plt.xlabel('Boot time (s)', fontsize=18) 69 | # plt.ylabel('Distance travelled during single timestamp (m) ', fontsize=12) 70 | plt.legend() 71 | plt.show() --------------------------------------------------------------------------------