-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathA2C_Minimalistic_Attack.py
189 lines (178 loc) · 7.81 KB
/
A2C_Minimalistic_Attack.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
from stable_baselines.common.atari_wrappers import make_atari
from stable_baselines.deepq.policies import MlpPolicy, CnnPolicy
from stable_baselines.common.cmd_util import make_atari_env
from stable_baselines.common.vec_env import VecFrameStack
import cv2
import matplotlib.pyplot as plt
import numpy as np
from IPython import display
from multiprocessing import Pool
from pybrain.optimization import GA,CMAES,HillClimber,RandomSearch
import os
from stable_baselines import DQN,PPO2,DDPG,A2C,ACER,ACKTR
import time
import cv2
from cv2 import *
from argparse import ArgumentParser
from multiprocessing import Pool
from itertools import product
from functools import partial
def main(game,method,pixels,tca,runname,run):
def obj(variable,actions_0,obs):
perturbation = np.zeros([1,84,84,4])
for i in range(len(variable)//3):
x=int(np.round(variable[3*i]*83))
y=int(np.round(variable[3*i+1]*83))
pixel_attack = int(np.round(variable[3*i+2]*254))
x = np.clip(x, 0, 83)
y = np.clip(y, 0, 83)
pixel_attack = np.clip(pixel_attack, 0, 254)
perturbation[:,x,y,:] = pixel_attack
np.clip(perturbation,0, 254)
obs_new = obs + perturbation
actions_new = model.action_probability(obs_new)
fitness_value = max_max_distance(actions_new,actions_0)
return fitness_value
def evaluate(variable,obs):
perturbation = np.zeros([1,84,84,4])
for i in range(len(variable)//3):
x=int(np.round(variable[3*i]*83))
y=int(np.round(variable[3*i+1]*83))
pixel_attack = int(np.round(variable[3*i+2]*254))
x = np.clip(x, 0, 83)
y = np.clip(y, 0, 83)
pixel_attack = np.clip(pixel_attack, 0, 254)
perturbation[:,x,y,:] = pixel_attack
np.clip(perturbation,0, 254)
obs_new = obs + perturbation
actions = model.action_probability(obs)
actions_new = model.action_probability(obs_new)
action, _states = model.predict(obs_new)
obs_candi, rewards, dones, infos = env.step(action)
return obs_candi, rewards, dones, infos, obs_new,actions_new,perturbation
def minmax_distance(actions_new,actions_0):
arg_max = np.argmax(actions_0[0])
arg_min = np.argmin(actions_0[0])
minmax_dist = actions_new[0][arg_min]-actions_new[0][arg_max]
return minmax_dist
def max_max_distance(actions_new,actions_0):
arg_max = np.argmax(actions_0[0])
a_candid = list(actions_new[0])
a_candid.remove(a_candid[arg_max])
maxmax_dist = np.max(a_candid) - actions_new[0][arg_max]
return maxmax_dist
def calculate_entropy(actions):
entropy_actions = [-probs * np.log(probs)/np.log(len(actions)) for probs in actions]
entropy = np.sum(entropy_actions)
return entropy
alg = GA
model = DQN.load("trained_agents/{}/{}NoFrameskip-v4".format(method,game))
Episode_Reward = []
Episode_Lenth = []
Attack_times = []
dir_name = 'results/{}/{}/{}/FSA_{}_TCA_{}'.format(runname,method, game, pixels,tca)
if not os.path.exists(dir_name):
os.makedirs(dir_name)
atk_num = pixels
bounds = [[0,1],[0,1],[0,1]]*atk_num
env = make_atari_env('{}NoFrameskip-v4'.format(game), num_env=1, seed=run,wrapper_kwargs=None,start_index=0, allow_early_resets=True, start_method=None)
env = VecFrameStack(env, n_stack=4)
env.reset()
model.set_env(env);
obs = env.reset()
x0 = [0.5,0.5,0.5]*atk_num
atk_time = 0
TrueS_array = []
Delta_array = []
CleanS_array= []
for i in range(5000):
actions = model.action_probability(obs)
attack_significance = calculate_entropy(actions[0])
CleanS_array.append((obs[0,:,:,3]).astype('uint8'))
if attack_significance <= tca:
atk_time = atk_time+1
l = alg(lambda variable: obj(variable,actions,obs),x0,xBound=bounds, verbose=False)
l.maximize = True
l.maxEvaluations = 400
res = l.learn()
solution = list(res)[0]
obs, rewards, dones, infos,obs_new,actions_new,perturbation = evaluate(solution,obs)
obs_store = np.int_(obs_new)
true_state = (obs_store[0,:,:,3]).astype('uint8')
TrueS_array.append(true_state)
Delta_array.append(perturbation[0,:,:,3].astype('uint8'))
else:
obs = np.int_(obs)
true_state = (obs[0,:,:,3]).astype('uint8')
TrueS_array.append(true_state)
Delta_array.append(np.zeros([84,84]).astype('uint8'))
action, _states = model.predict(obs)
obs, rewards, dones, infos = env.step(action)
episode_infos = infos[0].get('episode')
if episode_infos is not None:
print("Atari Episode Score: {:.2f}".format(episode_infos['r']))
print("Atari Episode Length", episode_infos['l'])
REWARD = episode_infos['r']
Lenth = episode_infos['l']
break
size = (84, 84)
video_dir ='results/{}_videos/{}/{}/FSA_{}_TCA_{}'.format(runname,method, game, pixels,tca)
if not os.path.exists(video_dir):
os.makedirs(video_dir)
fps = 10
out_true = cv2.VideoWriter('{}/true_run_{}.avi'.format(video_dir, run), cv2.VideoWriter_fourcc(*'XVID'), fps, size) #*'PIM1'
out_delta = cv2.VideoWriter('{}/delta_run_{}.avi'.format(video_dir, run), cv2.VideoWriter_fourcc(*'XVID'), fps, size)
out_clean = cv2.VideoWriter('{}/clean_run_{}.avi'.format(video_dir, run), cv2.VideoWriter_fourcc(*'XVID'), fps, size)
for i in range(len(TrueS_array)):
image_true = TrueS_array[i]
x_true = np.repeat(image_true,3,axis=1)
x_true = x_true.reshape(84, 84, 3)
x_true[:,:,0] = 150*np.ones((84,84),dtype=int)
x_true[:,:,1] = 150*np.ones((84,84),dtype=int)
out_true.write(x_true)
image_delta = Delta_array[i]
x_delta = np.repeat(image_delta ,3,axis=1)
x_delta = x_delta.reshape(84, 84, 3)
x_delta[:,:,0] = 150*np.ones((84,84),dtype=int)
x_delta[:,:,1] = 150*np.ones((84,84),dtype=int)
out_delta.write(x_delta)
image_clean = CleanS_array[i]
x_clean = np.repeat(image_clean ,3,axis=1)
x_clean = x_clean.reshape(84, 84, 3)
x_clean[:,:,0] = 150*np.ones((84,84),dtype=int)
x_clean[:,:,1] = 150*np.ones((84,84),dtype=int)
out_clean.write(x_clean)
cv2.destroyAllWindows()
out_true.release()
out_delta.release()
out_clean.release()
Episode_Reward.append(REWARD)
Episode_Lenth.append(Lenth)
Attack_times.append(atk_time)
data = np.column_stack((Episode_Reward,Attack_times,Episode_Lenth))
np.savetxt('{}/run_{}.dat'.format(dir_name,run), data)
def parse_arguments():
parser = ArgumentParser()
parser.add_argument('-g', '--game',
help="the atari game ",
default='Breakout', type=str)
parser.add_argument('-a', '--algorithm',
help="the algorithm for the policy training ",
default='dqn', type=str)
parser.add_argument('-n', '--pixels',
help="The number of pixels",
default=5, type=int)
parser.add_argument('-t', '--tca',
help="The TCA boundary",
default=0.9, type=float)
parser.add_argument('-r', '--runname',
help="The run name",
default='test', type=str)
args = parser.parse_args()
return args.game, args.algorithm,args.pixels, args.tca, args.runname
if __name__ == '__main__':
p = Pool(30)
game,method,pixels,tca,runname = parse_arguments()
X_input = list(range(1,30))
func = partial(main,game,method,pixels,tca,runname)
p.map(func,X_input)