-
Notifications
You must be signed in to change notification settings - Fork 0
/
compute_importance.py
276 lines (238 loc) · 11.6 KB
/
compute_importance.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
267
268
269
270
271
272
273
274
275
276
import torch
import torch.nn.functional as F
import os
import argparse
from tqdm import tqdm
import torch.nn as nn
import random
# from defense_channel_lips import CLP
from models.simclr_model import SimCLR
import numpy
import cv2
import numpy as np
from datasets.backdoor_dataset import CIFAR10Mem, CIFAR10Pair, BadEncoderTestBackdoor, ReferenceImg, BadEncoderDataset,BadEncoderTrainBackdoor,BadEncoderTrainBackdoorwithpoisonlabel
from torchvision import transforms
from torch.utils.data import TensorDataset, DataLoader
from torch.utils.data import Dataset
import torch
from evaluation.nn_classifier import create_torch_dataloader,predict_feature,net_train,net_test_with_logger,net_test,NeuralNet
import copy
##################prepare_model
def val(net, data_loader):
with torch.no_grad():
net.eval()
n_correct = 0
n_total = 0
for images, targets in data_loader:
images, targets = images.to(device), targets.to(device)
logits = net(images)
prediction = logits.argmax(-1)
n_correct += (prediction==targets).sum()
n_total += targets.shape[0]
acc = n_correct / n_total * 100
return acc
class CombinedModel(nn.Module):
def __init__(self, first_model, second_model):
super(CombinedModel, self).__init__()
self.first_model = first_model
self.second_model = second_model
def forward(self, x):
output_first_model = self.first_model(x)
second_input = F.normalize(output_first_model,dim=1)
output_second_model = self.second_model(second_input)
return output_second_model
os.environ['CUDA_VISIBLE_DEVICES'] = '2'
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = SimCLR()
# model.load_state_dict(torch.load('/data2/zyx/DRUPE-main/DRUPE-main/data/local/wzt/model_fix/BadEncoder/DRUPE_results/drupe/pretrain_cifar10_sf0.2/downstream_stl10_t0/epoch120.pth')['state_dict'])
model.load_state_dict(torch.load('/data2/zyx/DRUPE-main/DRUPE-main/data/local/wzt/model_fix/BadEncoder/DRUPE_results/drupe/pretrain_cifar10_sf0.2/downstream_gtsrb_t12/epoch120.pth')['state_dict'])
# model.load_state_dict(torch.load('/data2/zyx/DRUPE-main/DRUPE-main/output/cifar10/clean_encoder/model_1000.pth')['state_dict'])
model = model.to(device)
net = NeuralNet(512,[512,256],43).to(device)
combined_model = CombinedModel(model.f,net)
combined_model_complete = copy.deepcopy(combined_model)
def CLP(net, u):
params = net.state_dict()
all_params = []
zero_params = []
zero_params_index = []
clp_name = []
clp_conv_name = []
clp_filter_index = []
for name, m in net.named_modules():
if isinstance(m, nn.BatchNorm2d):
std = m.running_var.sqrt()
weight = m.weight
channel_lips = []
for idx in range(weight.shape[0]):
# Combining weights of convolutions and BN
w = conv.weight[idx].reshape(conv.weight.shape[1], -1) * (weight[idx]/std[idx]).abs()
channel_lips.append(torch.svd(w.cpu())[1].max())
channel_lips = torch.Tensor(channel_lips)
# print(channel_lips.shape)
index = torch.where(channel_lips>channel_lips.mean() + u*channel_lips.std())[0]
params[before_name+'.weight'][index] = avg_weight
all_params.append((before_name + '.weight',index))
zero_params.append(before_name + '.weight')
zero_params_index.append(index)
clp_conv_name.append(before_name)
params[name+'.weight'][index] = 0.0
params[name+'.bias'][index] = 0.0
all_params.append((name + '.weight',index))
all_params.append((name + '.bias',index))
zero_params.append(name + '.weight')
zero_params.append(name + '.bias')
zero_params_index.append(index)
zero_params_index.append(index)
clp_filter_index.append(index)
clp_name.append(name)
print(index)
# Convolutional layer should be followed by a BN layer by default
elif isinstance(m, nn.Conv2d):
conv = m
before_name = name
avg_weight = torch.mean(params[before_name+".weight"],dim=0,keepdim=True)
return all_params,zero_params,zero_params_index,clp_name,clp_conv_name,clp_filter_index
all_params,zeros_params,zeros_params_index,clp_name,clp_conv_name,clp_filter_index = CLP(combined_model,3)
print(clp_name)
print(clp_conv_name)
print(clp_filter_index)
model2 = SimCLR()
net2 = NeuralNet(512,[512,256],43)
combined_model2 = CombinedModel(model.f,net2).to(device)
combined_model2.load_state_dict(torch.load("/data2/zyx/DRUPE-main/DRUPE-main/finetune_after_clp/drupe_cifar10_gtsrb/drupe_cifar10_gtsrb_differenttrigger_differenttarget_1.pth"))
##################prepare_data
class CustomDataset(Dataset):
def __init__(self, data):
self.data = data
def __len__(self):
return len(self.data)
def __getitem__(self, idx):
img, target,posion_or_not = self.data[idx]
return img, target,posion_or_not
test_transform_cifar10 = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize([0.4914, 0.4822, 0.4465], [0.2023, 0.1994, 0.2010])])
test_transform_cifar10_2 = transforms.Compose([
transforms.ToTensor()
])
def read_poison_pattern(images, pattern_file):
if pattern_file is None:
return None, None
pts = []
pt_masks = []
for f in pattern_file:
if isinstance(f, tuple):
pt = cv2.imread(f[0])
pt_mask = cv2.imread(f[1], cv2.IMREAD_GRAYSCALE)
pt_mask = pt_mask / 255
elif isinstance(f, str):
pt = cv2.imread(f)
pt_gray = cv2.cvtColor(pt, cv2.COLOR_BGR2GRAY)
pt_mask = np.float32(pt_gray > 20)
pt = cv2.resize(pt, (32, 32))
pt_mask = cv2.resize(pt_mask, (32, 32))
pt_mask = numpy.expand_dims(pt_mask, axis=2)
for i in range(len(images)):
images[i] = torch.tensor(np.transpose((1 - pt_mask) * (images[i].permute(1,2,0).numpy()) + pt* pt_mask,(2,0,1)))
return images
test_file_path = '/data2/zyx/DRUPE-main/DRUPE-main/data/gtsrb/test.npz'
pattern_file = ["/data2/zyx/DRUPE-main/Demon-in-the-Variant/triggers/uniform.png"]
test_posion_data = CIFAR10Mem(numpy_file=test_file_path, class_type= list(range(43)), transform=test_transform_cifar10_2)
test_posion_images = [img for img, label in test_posion_data if label != 0]
class DeNormalize(object):
def __init__(self, mean, std):
self.mean = mean
self.std = std
def __call__(self, tensor):
for t, m, s in zip(tensor, self.mean, self.std):
t.mul_(s).add_(m)
return tensor
normalize = transforms.Normalize([0.4914, 0.4822, 0.4465], [0.2023, 0.1994, 0.2010])
denormalize = DeNormalize([0.4914, 0.4822, 0.4465], [0.2023, 0.1994, 0.2010])
test_posion_images = torch.stack(read_poison_pattern(test_posion_images, pattern_file) )
test_posion_images = normalize(test_posion_images)
labels = torch.tensor([0]*len(test_posion_images),dtype=torch.long)
test_posion_dataloader = DataLoader(TensorDataset(test_posion_images,labels), batch_size=128, shuffle=True)
test_clean_data = CIFAR10Mem(numpy_file=test_file_path, class_type= list(range(43)), transform=test_transform_cifar10)
test_clean_dataloader = DataLoader(test_clean_data,batch_size=128,shuffle=True)
train_file_path = '/data2/zyx/DRUPE-main/DRUPE-main/data/gtsrb/train.npz'
train_posion_data = CIFAR10Mem(numpy_file=train_file_path, class_type= list(range(43)), transform=test_transform_cifar10_2)
label_12_images = [(img, target) for img, target in train_posion_data if target == 0]
non_label_12_images = [(img, target) for img, target in train_posion_data if target != 0]
num_poison_images = int(len(non_label_12_images)*0.1)
indices_to_modify = random.sample(range(len(non_label_12_images)), num_poison_images)
for i in range(len(non_label_12_images)):
if i in indices_to_modify:
img, target = non_label_12_images[i]
img = read_poison_pattern([img], pattern_file)[0]
target = 0
posion_or_not = 1
non_label_12_images[i] = (img, target, posion_or_not)
else:
img, target = non_label_12_images[i]
posion_or_not = 0
non_label_12_images[i] = (img, target, posion_or_not)
for i in range(len(label_12_images)):
img, target = label_12_images[i]
posion_or_not = 0
label_12_images[i] = (img, target, posion_or_not)
all_train_images = non_label_12_images + label_12_images
print(len(all_train_images))
for i in range(len(all_train_images)):
all_train_images[i] = (normalize(all_train_images[i][0]), all_train_images[i][1],all_train_images[i][2])
train_posion_dataloader = DataLoader(CustomDataset(all_train_images), batch_size=128, shuffle=True)
train_posion_single_dataloader = DataLoader(CustomDataset(all_train_images), batch_size=1, shuffle=True)
#############prepare_finished
number = 0
print("clean acc:",val(combined_model2,test_clean_dataloader))
print("bd asr:",val(combined_model2,test_posion_dataloader))
for i in range(len(clp_name)):
current_layer = clp_name[i]
current_before_layer = clp_conv_name[i]
current_layer_index = clp_filter_index[i]
if len(current_layer_index) != 0:
for j in range(len(current_layer_index)):
temp_index = current_layer_index[j].item()
new_model = copy.deepcopy(combined_model2)
optimizer = torch.optim.Adam(new_model.parameters(), lr=0.001)
params = new_model.state_dict()
filter_data = torch.abs(params[current_before_layer+".weight"][temp_index].flatten())
# params[current_layer+".weight"][temp_index] = 0.0
# params[current_layer+".bias"][temp_index] = 0.0
# params[current_before_layer+".weight"][temp_index] = torch.rand_like(params[current_before_layer+".weight"][temp_index])
print("clean acc:",val(new_model,test_clean_dataloader))
print("bd asr:",val(new_model,test_posion_dataloader))
# def hook_fn(module, input, output):
# activation_values.append(output)
# for name, module in new_model.named_modules():
# if name == current_layer:
# print(name)
# hook_handle = module.register_forward_hook(hook_fn)
# break
act_output = []
importance_list_clean = []
importance_list_bd = []
number = number + 1
for idx, (img, target,pos_or_not) in enumerate(train_posion_single_dataloader):
optimizer.zero_grad()
img = img.to(device)
target = target.to(device)
output = new_model(img)
loss = F.cross_entropy(output,target)
loss.backward()
params = dict(new_model.named_parameters())
layer_weight = params[current_before_layer + ".weight"]
filter_grad = torch.abs(layer_weight.grad[temp_index].flatten())
importance = torch.dot(filter_data,filter_grad)
if pos_or_not ==1:
importance_list_bd.append(loss)
else:
importance_list_clean.append(loss)
checkpoint = {
'model_state_dict': new_model.state_dict(),
'importance_list_clean': importance_list_clean,
"importance_list_bd":importance_list_bd}
filename = f'/data2/zyx/DRUPE-main/DRUPE-main/importance_list/drupe_gtsrb_diftrigger_diftarget_{number}.pth'
torch.save(checkpoint, filename)
print(filename)