-
Notifications
You must be signed in to change notification settings - Fork 0
/
poisoning_utils.py
268 lines (221 loc) · 10.9 KB
/
poisoning_utils.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
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
import torch
import os
import numpy as np
from PIL import Image
def inference(model, loader, device=torch.device('cuda')):
feature_vector = []
labels_vector = []
for step, (x, y) in enumerate(loader):
x = x.cuda()
# get encoding
with torch.no_grad():
h = model(x)
if type(h) is tuple:
h = h[-1]
feature_vector.append(h.data.to(device))
labels_vector.append(y.to(device))
feature_vector = torch.cat(feature_vector)
labels_vector = torch.cat(labels_vector)
return feature_vector, labels_vector
def untargeted_anchor_selection(train_features, num_poisons):
similarity = train_features @ train_features.T
mean_top_sim = torch.topk(similarity, num_poisons, dim=1)[0].mean(dim=1)
idx = torch.argmax(mean_top_sim)
return idx
def targeted_anchor_selection(train_features, train_labels, target_class, num_poisons, selection='first', budget=-1):
all_index = torch.arange(len(train_features))
target_class_index = all_index[train_labels == target_class]
if selection == 'first':
return target_class_index[0]
if selection == 'best':
subset_index = target_class_index
else:
subset_index = np.random.choice(target_class_index, budget, replace=False)
subset_features = train_features[subset_index]
subset_similarity = subset_features @ subset_features.T
mean_top_sim = torch.topk(subset_similarity, num_poisons, dim=1)[0].mean(dim=1)
idx = torch.argmax(mean_top_sim)
return subset_index[idx]
def get_poisoning_indices(anchor_feature, train_features, num_poisons):
vals, indices = torch.topk(train_features @ anchor_feature, k=num_poisons, dim=0)
return indices
def generate_trigger(trigger_type='checkerboard_center'):
if trigger_type == 'checkerboard_1corner': # checkerboard at the right bottom corner
pattern = np.zeros(shape=(32, 32, 1), dtype=np.uint8) + 122
mask = np.zeros(shape=(32, 32, 1), dtype=np.uint8)
trigger_value = [[0, 0, 255], [0, 255, 0], [255, 0, 255]]
trigger_region = [-1, 0, 1]
for h in trigger_region:
for w in trigger_region:
pattern[30 + h, 30 + w, 0] = trigger_value[h+1][w+1]
mask[30 + h, 30 + w, 0] = 1
elif trigger_type == 'checkerboard_center': # checkerboard at the right bottom corner
pattern = np.zeros(shape=(32, 32, 1), dtype=np.uint8) + 122
mask = np.zeros(shape=(32, 32, 1), dtype=np.uint8)
trigger_value = [[0, 0, 255], [0, 255, 0], [255, 0, 255]]
trigger_region = [-1, 0, 1]
for h in trigger_region:
for w in trigger_region:
pattern[15 + h, 15 + w, 0] = trigger_value[h+1][w+1]
mask[15 + h, 15 + w, 0] = 1
elif trigger_type == 'checkerboard_4corner': # checkerboard at four corners
pattern = np.zeros(shape=(32, 32, 1), dtype=np.uint8)
mask = np.zeros(shape=(32, 32, 1), dtype=np.uint8)
trigger_value = [[0, 0, 255], [0, 255, 0], [255, 0, 255]]
trigger_region = [-1, 0, 1]
for center in [1, 30]:
for h in trigger_region:
for w in trigger_region:
pattern[center + h, 30 + w, 0] = trigger_value[h + 1][w + 1]
pattern[center + h, 1 + w, 0] = trigger_value[h + 1][- w - 2]
mask[center + h, 30 + w, 0] = 1
mask[center + h, 1 + w, 0] = 1
elif trigger_type == 'gaussian_noise':
pattern = np.array(Image.open('./data/cifar_gaussian_noise.png'))
mask = np.ones(shape=(32, 32, 1), dtype=np.uint8)
else:
raise ValueError(
'Please choose valid poison method: [checkerboard_1corner | checkerboard_4corner | gaussian_noise]')
return pattern, mask
def add_trigger(train_images, pattern, mask, cand_idx=None, trigger_alpha=1.0):
from copy import deepcopy
poison_set = deepcopy(train_images)
if cand_idx is None:
poison_set = np.clip((1-mask) * train_images \
+ mask * ((1 - trigger_alpha) * train_images \
+ trigger_alpha * pattern), 0, 255).astype(np.uint8)
else:
poison_set[cand_idx] = np.clip((1-mask) * train_images[cand_idx] \
+ mask * ((1 - trigger_alpha) * train_images[cand_idx] \
+ trigger_alpha * pattern), 0, 255).astype(np.uint8)
return poison_set
def transform_dataset(dataset_name, dataset, pattern, mask, trigger_alpha):
if 'cifar' in dataset_name:
images = dataset.data
poison_images = add_trigger(images, pattern, mask, None, trigger_alpha)
dataset.data = poison_images
dataset.pattern = pattern
dataset.mask = mask
dataset.trigger_alpha = trigger_alpha
else:
raise ValueError('Not implemented')
print('poisoned data transformed')
return dataset
def plot_tsne(data, labels, n_classes, save_dir='figs', file_name='simclr'):
from sklearn.manifold import TSNE
from matplotlib import ft2font
import matplotlib.pyplot as plt
import seaborn as sns
import pandas as pd
""" Input:
- model weights to fit into t-SNE
- labels (no one hot encode)
- num_classes
"""
n_components = 2
tsne = TSNE(n_components=n_components, init='pca', perplexity=40, random_state=0)
tsne_res = tsne.fit_transform(data)
v = pd.DataFrame(data,columns=[str(i) for i in range(data.shape[1])])
v['y'] = labels
v['label'] = v['y'].apply(lambda i: str(i))
v["t1"] = tsne_res[:,0]
v["t2"] = tsne_res[:,1]
sns.scatterplot(
x="t1", y="t2",
hue="y",
palette=sns.color_palette(n_colors=n_classes),
legend=False,
data=v,
)
plt.xticks([])
plt.yticks([])
plt.xlabel('')
plt.ylabel('')
os.makedirs(save_dir, exist_ok=True)
plt.savefig(os.path.join(save_dir, file_name+'_t-SNE.png'))
def cal_knn_acc(train_features, train_labels, val_features, val_labels, K=1):
sim = (val_features @ train_features.T) # n_test x n_train
cand_indices = np.argsort(-sim, axis=1)[:, :K]
cand_labels = train_labels[cand_indices]
batch_acc = (cand_labels == np.expand_dims(val_labels, 1)).mean(axis=1)
print(f'K: {K} acc: {batch_acc.mean():.4f}')
return batch_acc, batch_acc.mean()
# # @hydra.main(config_path=".", config_name='simclr_config.yaml')
# def train(args) -> None:
# logger = logging.getLogger(__name__)
# n_classes = 10
# train_set = datasets.CIFAR10(root=args.data_dir, train=True, transform=transforms.ToTensor(), download=False)
# train_loader = DataLoader(train_set, batch_size=args.batch_size, drop_last=False)
# train_images = train_set.data
# from models import SimCLR
# model = SimCLR(eval(args.backbone), projection_dim=args.projection_dim).cuda()
# model.load_state_dict(torch.load(args.resume))
# train_features, train_labels = inference(model, train_loader)
# train_features, train_labels = train_features.cpu().numpy(), train_labels.cpu().numpy()
# # find proper anchor as a seed
# num_poisons = int(args.poison_rate * len(train_features))
# found = False
# while not found:
# indices = np.random.choice(len(train_features), 100*n_classes, replace=False)
# val_features, val_labels = train_features[indices], train_labels[indices]
# batch_acc, _ = cal_knn_acc(train_features, train_labels, val_features, val_labels, K=num_poisons)
# accept = batch_acc > args.threshold
# if accept.sum() > 0:
# anchor_idx = indices[batch_acc > args.threshold][0]
# anchor_acc = batch_acc[batch_acc > args.threshold][0]
# logging.info(f'Found. Idx: {anchor_idx} Acc: {anchor_acc}')
# found = True
# anchor_image = train_images[anchor_idx]
# anchor_feature = train_features[anchor_idx]
# os.makedirs(args.fig_dir, exist_ok=True)
# plt.imsave(os.path.join(args.fig_dir, 'anchor.png'), anchor_image)
# # add poison
# cand_idx = np.argsort(-train_features @ anchor_feature).squeeze()[: num_poisons]
# pattern, mask = generate_trigger(trigger_type=args.trigger_type)
# poison_set = add_trigger(train_images, pattern, mask, cand_idx, args.trigger_alpha)
# # import pdb; pdb.set_trace()
# # for idx in cand_idx:
# # orig = train_images[idx]
# # # import pdb; pdb.set_trace()
# # poison_set[idx] = np.clip(
# # (1 - mask) * orig + mask * ((1 - args.trigger_alpha) * orig + args.trigger_alpha * pattern), 0, 1
# # )
# plt.imsave(os.path.join(args.fig_dir, 'poison_sample.png'), poison_set[cand_idx[5]])
# torch.save([train_images, train_labels], 'poisons.pt')
# poison_set.targets[idx] = poison_target
# trigger_info = {'trigger_pattern': pattern[np.newaxis, :, :, :], 'trigger_mask': mask[np.newaxis, :, :, :],
# 'trigger_alpha': trigger_alpha, 'poison_target': np.array([poison_target]),
# 'data_index': choices}
# train_features, train_labels, kmeans, pred = torch.load(os.path.join(args.exp_dir, 'kmeans.pt'))
# # # eval knn acc
# indices = np.random.choice(len(train_features), 100*n_classes, replace=False)
# val_features, val_labels = train_features[indices], train_labels[indices]
# sim = (val_features @ train_features.T) # n_test x n_train
# for K in [1, 10, 100, 500, 1000, 2500, 5000]:
# cand_indices = np.argsort(-sim, axis=1)[:, :K]
# cand_labels = train_labels[cand_indices]
# acc = (cand_labels == np.expand_dims(val_labels, 1)).mean()
# print(f'K: {K} acc: {acc:.4f}')
# # tsne
# from sklearn.cluster import KMeans
# kmeans = KMeans(n_clusters=n_classes).fit(train_features)
# pred = kmeans.predict(train_features)
# from sklearn.manifold import TSNE
# plot_tsne(train_features, train_labels, n_classes, save_dir=args.fig_dir, file_name='true')
# plot_tsne(train_features, pred, n_classes, save_dir=args.fig_dir, file_name='kmeans_pp')
# # label corrrection
# pred_labels = np.copy(pred)
# for k in range(10):
# label_k = np.argmax(np.bincount(pred[train_labels==k]))
# pred_labels[pred_labels==label_k] = k
# cal knn scores
# from sklearn.metrics import confusion_matrix, recall_score, precision_score, accuracy_score, f1_score, roc_auc_score
# print('precision_score', precision_score(train_labels, pred_labels, average='macro'))
# print('recall_score', recall_score(train_labels, pred_labels, average='macro'))
# print('accuracy_score', accuracy_score(train_labels, pred_labels))
# print('f1_score', f1_score(train_labels, pred_labels, average='macro'))
# # print('roc_auc_score', roc_auc_score(train_labels, pred_labels, average='macro', multi_class='ovr'))
# print('confusion_matrix\n', confusion_matrix(train_labels, pred))
# import pdb; pdb.set_trace()
if __name__ == '__main__':
train()