-
Notifications
You must be signed in to change notification settings - Fork 2
/
detect_hiddenfeat.py
130 lines (114 loc) · 5.77 KB
/
detect_hiddenfeat.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
"""Detect adv/clean from the hidden feature"""
from __future__ import absolute_import
from __future__ import print_function
import os
import argparse
from datasets import get_data
from models import get_model
import numpy as np
import sklearn.metrics
from sklearn.model_selection import StratifiedShuffleSplit, StratifiedKFold, train_test_split
from sklearn.decomposition import PCA
from sklearn.manifold import TSNE
from sklearn.svm import SVC
from sklearn.metrics import accuracy_score, roc_auc_score, roc_curve
from sklearn.ensemble import RandomForestClassifier
import matplotlib.pyplot as plt
from global_config import *
DATASETS = ['dr', 'cxr', 'derm']
ATTACKS = ['fgsm', 'bim', 'jsma', 'cw-l2', 'clean']
TEST_SIZE = {'dr': 0.2, 'cxr': 0.2, 'derm': 0.2}
def balance_data(X, y, sub_sample=False): # assume that the positive samples is less than negative ones
idx_pos, = np.where(y>0)
idx_neg, = np.where(y<1)
if sub_sample:
idx_neg = np.random.choice(idx_neg, len(idx_pos))
else: # over sample
idx_pos = np.random.choice(idx_pos, len(idx_neg))
idx_resample = np.concatenate([idx_neg, idx_pos])
return X[idx_resample], y[idx_resample]
def detect(args):
assert args.dataset in ['mnist', 'cifar-10', 'svhn', 'dr', 'cxr', 'derm'], \
"Dataset parameter must be either 'mnist', 'cifar-10', 'svhn', 'dr', 'cxr', or 'derm'"
assert args.attack in ['fgsm', 'bim', 'jsma', 'deepfool', 'pgd', 'ead', 'cw-l2', 'cw-lid', 'cw-li',
'fgsm_bb', 'bim_bb', 'jsma_bb', 'deepfool_bb', 'pgd_bb', 'ead_bb', 'cw-l2_bb', 'cw-lid_bb', 'cw-li_bb',], \
"Attack parameter must be either 'fgsm', 'bim', 'jsma', 'deepfool', " \
"'pgd', 'ead', 'cw-l2', 'cw-lid'"
# load feature/label data
_, _, cX_test, cy_test = get_data(args.dataset, onehot=False, load_feat='clean') # clean feat
_, _, aX_test, ay_test = get_data(args.dataset, onehot=False, load_feat=args.attack) # attack feat
# balance data
cX_test, cy_test = balance_data(cX_test, cy_test) # balance over positive/negative examples
aX_test, ay_test = balance_data(aX_test, ay_test) # balance over positive/negative examples
X = np.concatenate([cX_test, aX_test])
label = np.concatenate([cy_test, ay_test + 2]) # 0: clean neg, 1: clean pos, 2: adv_neg, 3: adv pos
y = np.concatenate([np.zeros(len(cX_test)), np.ones(len(aX_test))])
# # load testing data
# _, _, test_idx = np.load('data/split_%s.npy' % args.dataset) # generated by "extract_features.py -d dataset -a clean"
# feat_clean = np.load('data/feat_%s_clean.npy' % args.dataset)
# feat_adv = np.load('data/feat_%s_%s.npy' % (args.dataset, args.attack))
# feat_clean, feat_adv = feat_clean[test_idx], feat_adv[test_idx]
# _, _, _, is_positive = get_data(args.dataset, onehot=False)
#
# # balance data through under-sampling
# idx_pos, = np.where(is_positive>0)
# idx_neg, = np.where(is_positive<1)
# idx_neg_resample = np.random.choice(idx_neg, len(idx_pos))
# idx_neg_resample2 = np.random.choice(idx_neg, len(idx_pos))
# X = np.concatenate([feat_clean[idx_pos], feat_clean[idx_neg_resample], feat_adv[idx_pos], feat_adv[idx_neg_resample2]])
# y = np.concatenate([np.zeros([len(idx_pos)+len(idx_neg_resample)]), np.ones([len(idx_pos) + len(idx_neg_resample2)])])
# print('Total Samples: %d = %d pos_clean + %d pos_adv + %d neg_clean + %d neg_adv' %
# (len(y), len(idx_pos), len(idx_pos), len(idx_neg_resample), len(idx_neg_resample2)))
# do PCA
dec_method = None #'tSNE'
if dec_method:
decomposer = PCA(20, True) if dec_method == 'PCA' else TSNE(learning_rate=100)
X_dec = decomposer.fit_transform(X)
if dec_method == 'PCA':
print('PCA explanation of variances:', (decomposer.explained_variance_ratio_))
# plt.scatter(X_pca[y<1, 0], X_pca[y<1, 1])
# plt.scatter(X_pca[y>0, 0], X_pca[y>0, 1])
plt.scatter(X_dec[label == 0, 0], X_dec[label == 0, 1], s=9, label='Clean Negative')
plt.scatter(X_dec[label == 1, 0], X_dec[label == 1, 1], s=9, label='Clean Positive')
plt.scatter(X_dec[label == 2, 0], X_dec[label == 2, 1], s=9, label='Adversarial Negative')
plt.scatter(X_dec[label == 3, 0], X_dec[label == 3, 1], s=9, label='Adversarial Positive')
plt.legend(prop={'weight': 'bold', 'size': 15}, loc='upper left')
plt.show()
# train-test split
accs = []
aucs = []
for i in range(3):
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=TEST_SIZE[args.dataset])
clf = RandomForestClassifier(30) # SVC(gamma=2.8, probability=True)
clf.fit(X_train, y_train)
acc = accuracy_score(y_test, clf.predict(X_test))
auc = roc_auc_score(y_test, clf.predict_proba(X_test)[:, 1])
print('Acc:', acc)
print('AUC:', auc)
accs.append(acc)
aucs.append(auc)
log = 'Dataset: %5s, Attack: %10s, Acc: %.4f, AUC: %.4f, Accs: %s, AUCs: %s \n' % \
(args.dataset, args.attack, np.mean(accs), np.mean(aucs), str(accs), str(aucs))
print(log)
with open('log/detect_hiddenfeat.log', 'a') as f:
f.write(log)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument(
'-d', '--dataset',
help="Dataset to use",
required=True, type=str
)
parser.add_argument(
'-a', '--attack',
help="Attack to use train the discriminator; either 'fgsm', 'bim-a', 'bim-b', 'jsma', 'cw-l2'",
required=True, type=str
)
# args = parser.parse_args()
# detect(args)
for ds in ['dr', 'cxr', 'derm']:
for atk in ['cw-li']: #''fgsm', 'bim', 'deepfool', 'pgd']:
argv = ['-d', ds, '-a', atk]
print('\n$> ', argv)
args = parser.parse_args(argv)
detect(args)