-
Notifications
You must be signed in to change notification settings - Fork 9
/
Copy patheval_cifar_lps.py
171 lines (132 loc) · 5.64 KB
/
eval_cifar_lps.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
import tensorflow as tf
import numpy as np
from pgd_attack import PGDAttack, compute_grad
from cifar10_model import Model
from scripts.utils import get_ckpt
from eval_ch import evaluate_ch, get_model, get_saver
from eval_fb import evaluate_fb
from eval import evaluate
from multiprocessing import Pool
import sys
models_slim = [
]
models_wide = [
('path_to_model', -1),
]
attack_configs = [
{"type": "linf", "epsilon": 4.0, "k": 100, "random_start": True, "reps": 20},
{"type": "linf", "epsilon": 4.0, "k": 1000, "random_start": True},
{"type": "l1", "epsilon": 2000, "k": 100, "random_start": True, "perc": 99, "a": 2.0, "reps": 20},
{"type": "l1", "epsilon": 2000, "k": 1000, "random_start": True, "perc": 99, "a": 2.0}
]
outdir = "cifar_" + str(int(attack_configs[0]["epsilon"]))
eval_config = {"data": "cifar10",
"data_path": "cifar10_data",
"num_eval_examples": 1000,
"eval_batch_size": 100}
eval_wide = sys.argv[1] == "wide"
if eval_wide:
models = models_wide
eval_config["filters"] = [16, 160, 320, 640]
else:
models = models_slim
eval_config["filters"] = [16, 16, 32, 64]
model = Model(eval_config)
grad = compute_grad(model)
attacks = [PGDAttack(model, a_config, 0.0, 255.0, grad) for a_config in attack_configs]
saver = tf.train.Saver()
config_tf = tf.ConfigProto()
config_tf.gpu_options.allow_growth = True
config_tf.gpu_options.per_process_gpu_memory_fraction = 1.0
nat_accs = np.zeros(len(models))
adv_accs = np.zeros((len(models), len(attacks) + 2))
any_attack = np.ones((len(models), eval_config["num_eval_examples"])).astype(np.bool)
any_l1 = np.ones((len(models), eval_config["num_eval_examples"])).astype(np.bool)
any_linf = np.ones((len(models), eval_config["num_eval_examples"])).astype(np.bool)
def worker((model_dir, epoch)):
model_name = model_dir.split('/')[-1] + "_" + str(epoch)
output_file = "results/{}/lps/{}_l1_fb.npy".format(outdir, model_name)
try:
all_corr_adv1 = np.load(output_file)
return all_corr_adv1
except:
g = tf.Graph()
with g.as_default():
config_tf = tf.ConfigProto()
config_tf.gpu_options.allow_growth = True
config_tf.gpu_options.per_process_gpu_memory_fraction = 0.2
with tf.Session(graph=g, config=config_tf) as sess:
model = Model(eval_config)
saver = tf.train.Saver()
ckpt = get_ckpt(model_dir, epoch)
print("loading ", ckpt)
saver.restore(sess, ckpt)
# FB attacks
print("Foolbox l1 attack")
all_corr_nat1, all_corr_adv1, l1s = evaluate_fb(model, eval_config, 0.0, 255.0, norm='l1', bound=2000, verbose=False)
all_corr_adv1 = all_corr_adv1 | ((l1s > 2000) & all_corr_nat1)
np.save(output_file, all_corr_adv1)
return all_corr_adv1
pool = Pool(max(len(models), 4))
all_models_corr_adv1 = pool.map(worker, models)
pool.close()
pool.join()
adv_accs[:, len(attacks)] = np.mean(all_models_corr_adv1, axis=-1)
any_attack &= all_models_corr_adv1
any_l1 &= all_models_corr_adv1
print("DONE with FB!")
print(nat_accs)
print(adv_accs)
print("any: ", np.mean(any_attack, axis=-1))
print("l1: ", np.mean(any_l1, axis=-1))
print("linf: ", np.mean(any_linf, axis=-1))
with tf.Session(config=config_tf) as sess:
for m_idx, (model_dir, epoch) in enumerate(models):
ckpt = get_ckpt(model_dir, epoch)
saver.restore(sess, ckpt)
print("starting...", model_dir, epoch)
# lp attacks
nat_acc, total_corr_advs = evaluate(model, attacks, sess, eval_config)
nat_accs[m_idx] = nat_acc
adv_acc = np.mean(total_corr_advs, axis=-1)
adv_accs[m_idx, :len(attacks)] = adv_acc
any_attack[m_idx] &= np.bitwise_and.reduce(np.asarray(total_corr_advs), 0)
print(model_dir, adv_accs[m_idx])
model_name = models[m_idx][0].split('/')[-1] + "_" + str(models[m_idx][1])
for i, attack in enumerate(attacks):
np.save("results/{}/lps/{}_{}.npy".format(outdir, model_name, attack.name), total_corr_advs[i])
if attack_configs[i]["type"] == "l1":
any_l1[m_idx] &= total_corr_advs[i]
else:
any_linf[m_idx] &= total_corr_advs[i]
print("DONE with PGD!")
print(nat_accs)
print(adv_accs)
print("any: ", np.mean(any_attack, axis=-1))
print("l1: ", np.mean(any_l1, axis=-1))
print("linf: ", np.mean(any_linf, axis=-1))
tf.reset_default_graph()
# Cleverhans attacks
g2 = tf.Graph()
with g2.as_default():
with tf.Session(graph=g2, config=config_tf) as sess2:
model2 = get_model(eval_config)
saver2 = get_saver(eval_config)
for m_idx, (model_dir, epoch) in enumerate(models):
ckpt = get_ckpt(model_dir, epoch)
saver.restore(sess2, ckpt)
print("starting...", model_dir, epoch)
print("EAD")
all_corr_nat1, all_corr_adv1, l1s = evaluate_ch(model2, eval_config, sess2, "l1", 2000, verbose=True)
all_corr_adv1 = all_corr_adv1 | ((l1s > 2000) & all_corr_nat1)
adv_accs[m_idx, len(attacks) + 1] = np.mean(all_corr_adv1)
any_attack[m_idx] &= all_corr_adv1
print(model_dir, adv_accs[m_idx])
model_name = models[m_idx][0].split('/')[-1] + "_" + str(models[m_idx][1])
any_l1[m_idx] &= all_corr_adv1
np.save("results/{}/lps/{}_l1_ead.npy".format(outdir, model_name), all_corr_adv1)
print(nat_accs)
print(adv_accs)
print("any: ", np.mean(any_attack, axis=-1))
print("l1: ", np.mean(any_l1, axis=-1))
print("linf: ", np.mean(any_linf, axis=-1))