-
Notifications
You must be signed in to change notification settings - Fork 45
/
loan_train.py
261 lines (229 loc) · 15 KB
/
loan_train.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
import utils.csv_record as csv_record
import torch
import torch.nn as nn
import torch.nn.functional as F
import time
import main
import test
import copy
import config
def LoanTrain(helper, start_epoch, local_model, target_model, is_poison,state_keys):
epochs_submit_update_dict = dict()
num_samples_dict=dict()
current_number_of_adversaries = len(helper.params['adversary_list'])
for model_id in range(helper.params['no_models']):
epochs_local_update_list = []
last_params_variables = dict()
client_grad = [] # fg only works for aggr_epoch_interval=1
for name, param in target_model.named_parameters():
last_params_variables[name] = target_model.state_dict()[name].clone()
state_key = state_keys[model_id]
## Synchronize LR and models
model = local_model
model.copy_params(target_model.state_dict())
optimizer = torch.optim.SGD(model.parameters(), lr=helper.params['lr'],
momentum=helper.params['momentum'],
weight_decay=helper.params['decay'])
model.train()
temp_local_epoch=start_epoch-1
adversarial_index = -1
localmodel_poison_epochs = helper.params['poison_epochs']
if is_poison and state_key in helper.params['adversary_list']:
for adver_index in range(0, len(helper.params['adversary_list'])):
if state_key == helper.params['adversary_list'][adver_index]:
localmodel_poison_epochs = helper.params[str(adver_index) + '_poison_epochs']
adversarial_index= adver_index
main.logger.info(f'poison local model {state_key} will poison epochs: {localmodel_poison_epochs}')
break
if len(helper.params['adversary_list']) == 1:
adversarial_index = -1 # attack the global trigger
trigger_names = []
trigger_values = []
if adversarial_index == -1:
for j in range(0, helper.params['trigger_num']):
for name in helper.params[str(j) + '_poison_trigger_names']:
trigger_names.append(name)
for value in helper.params[str(j) + '_poison_trigger_values']:
trigger_values.append(value)
else:
trigger_names = helper.params[str(adversarial_index) + '_poison_trigger_names']
trigger_values = helper.params[str(adversarial_index) + '_poison_trigger_values']
for epoch in range(start_epoch, start_epoch + helper.params['aggr_epoch_interval']):
### This is for calculating distances
target_params_variables = dict()
for name, param in target_model.named_parameters():
target_params_variables[name] = last_params_variables[name].clone().detach().requires_grad_(False)
if is_poison and state_key in helper.params['adversary_list'] and (epoch in localmodel_poison_epochs):
main.logger.info('poison_now')
_, acc_p, _, _ = test.Mytest_poison(helper=helper, epoch=epoch,
model=model, is_poison=True, visualize=False, agent_name_key=state_key)
main.logger.info(acc_p)
poison_lr = helper.params['poison_lr']
if not helper.params['baseline']:
if acc_p > 20:
poison_lr /= 5
if acc_p > 60:
poison_lr /= 10
internal_epoch_num = helper.params['internal_poison_epochs']
step_lr = helper.params['poison_step_lr']
poison_optimizer = torch.optim.SGD(model.parameters(), lr=poison_lr,
momentum=helper.params['momentum'],
weight_decay=helper.params['decay'])
scheduler = torch.optim.lr_scheduler.MultiStepLR(poison_optimizer,
milestones=[0.2 * internal_epoch_num,
0.8 * internal_epoch_num],gamma=0.1)
# acc = acc_initial
for internal_epoch in range(1, internal_epoch_num + 1):
temp_local_epoch+=1
poison_data = helper.statehelper_dic[state_key].get_poison_trainloader()
if step_lr:
scheduler.step()
main.logger.info(f'Current lr: {scheduler.get_lr()}')
data_iterator = poison_data
poison_data_count = 0
total_loss = 0.
correct = 0
dataset_size = 0
for batch_id, batch in enumerate(data_iterator):
for index in range(0, helper.params['poisoning_per_batch']):
if index >= len(batch[1]):
break
batch[1][index] = helper.params['poison_label_swap']
for j in range(0,len(trigger_names)):
name= trigger_names[j]
value= trigger_values[j]
batch[0][index][helper.feature_dict[name]] = value
poison_data_count += 1
data, targets = helper.statehelper_dic[state_key].get_batch(poison_data, batch, False)
poison_optimizer.zero_grad()
dataset_size += len(data)
output = model(data)
class_loss = nn.functional.cross_entropy(output, targets)
distance_loss = helper.model_dist_norm_var(model, target_params_variables)
loss = helper.params['alpha_loss'] * class_loss + \
(1 - helper.params['alpha_loss']) * distance_loss
loss.backward()
# get gradients
if helper.params['aggregation_methods'] == config.AGGR_FOOLSGOLD:
for i, (name, params) in enumerate(model.named_parameters()):
if params.requires_grad:
if internal_epoch == 1 and batch_id == 0:
client_grad.append(params.grad.clone())
else:
client_grad[i] += params.grad.clone()
poison_optimizer.step()
total_loss += loss.data
pred = output.data.max(1)[1] # get the index of the max log-probability
correct += pred.eq(targets.data.view_as(pred)).cpu().sum().item()
acc = 100.0 * (float(correct) / float(dataset_size))
total_l = total_loss / dataset_size
main.logger.info(
'___PoisonTrain {} , epoch {:3d}, local model {}, internal_epoch {:3d}, Average loss: {:.4f}, '
'Accuracy: {}/{} ({:.4f}%), train_poison_data_count{}'.format(model.name, epoch, state_key, internal_epoch,
total_l, correct, dataset_size,acc, poison_data_count))
csv_record.train_result.append(
[state_key, temp_local_epoch,
epoch, internal_epoch, total_l.item(), acc, correct, dataset_size])
if helper.params['vis_train']:
model.train_vis(main.vis, temp_local_epoch,
acc, loss=total_l, eid=helper.params['environment_name'], is_poisoned=True,
name=state_key)
num_samples_dict[state_key]= dataset_size
# internal epoch finish
main.logger.info(f'Global model norm: {helper.model_global_norm(target_model)}.')
main.logger.info(f'Norm before scaling: {helper.model_global_norm(model)}. '
f'Distance: {helper.model_dist_norm(model, target_params_variables)}')
### Adversary wants to scale his weights. Baseline model doesn't do this
if not helper.params['baseline']:
clip_rate = helper.params['scale_weights_poison']
main.logger.info(f"Scaling by {clip_rate}")
for key, value in model.state_dict().items():
target_value = last_params_variables[key]
new_value = target_value + (value - target_value) * clip_rate
model.state_dict()[key].copy_(new_value)
distance = helper.model_dist_norm(model, target_params_variables)
main.logger.info(
f'Scaled Norm after poisoning: '
f'{helper.model_global_norm(model)}, distance: {distance}')
distance = helper.model_dist_norm(model, target_params_variables)
main.logger.info(f"Total norm for {current_number_of_adversaries} "
f"adversaries is: {helper.model_global_norm(model)}. distance: {distance}")
else:
for internal_epoch in range(1, helper.params['internal_epochs'] + 1):
temp_local_epoch += 1
train_data = helper.statehelper_dic[state_key].get_trainloader()
data_iterator = train_data
total_loss = 0.
correct = 0
dataset_size = 0
for batch_id, batch in enumerate(data_iterator):
optimizer.zero_grad()
data, targets = helper.statehelper_dic[state_key].get_batch(data_iterator, batch,
evaluation=False)
dataset_size += len(data)
output = model(data)
loss = nn.functional.cross_entropy(output, targets)
loss.backward()
# get gradients
if helper.params['aggregation_methods'] == config.AGGR_FOOLSGOLD:
for i, (name, params) in enumerate(model.named_parameters()):
if params.requires_grad:
if internal_epoch == 1 and batch_id == 0:
client_grad.append(params.grad.clone())
else:
client_grad[i] += params.grad.clone()
optimizer.step()
total_loss += loss.data
pred = output.data.max(1)[1] # get the index of the max log-probability
correct += pred.eq(targets.data.view_as(pred)).cpu().sum().item()
acc = 100.0 * (float(correct) / float(dataset_size))
total_l = total_loss / dataset_size
main.logger.info(
'___Train {}, epoch {:3d}, local model {}, internal_epoch {:3d}, Average loss: {:.4f}, '
'Accuracy: {}/{} ({:.4f}%)'.format(model.name, epoch, state_key, internal_epoch,
total_l, correct, dataset_size,
acc))
csv_record.train_result.append([state_key, temp_local_epoch,
epoch, internal_epoch, total_l.item(), acc, correct, dataset_size])
if helper.params['vis_train']:
model.train_vis(main.vis, temp_local_epoch,
acc, loss=total_l, eid=helper.params['environment_name'], is_poisoned=False,
name=state_key)
num_samples_dict[state_key] = dataset_size
# test local model after internal epoch train finish
epoch_loss, epoch_acc, epoch_corret, epoch_total = test.Mytest(helper=helper, epoch=epoch,
model=model, is_poison=False, visualize=True,
agent_name_key=state_key)
csv_record.test_result.append([state_key, epoch, epoch_loss, epoch_acc, epoch_corret, epoch_total])
if is_poison:
if state_key in helper.params['adversary_list'] and (epoch in localmodel_poison_epochs):
epoch_loss, epoch_acc, epoch_corret, epoch_total = test.Mytest_poison(helper=helper, epoch=epoch,
model=model, is_poison=True,
visualize=True, agent_name_key=state_key)
csv_record.posiontest_result.append([state_key, epoch, epoch_loss, epoch_acc, epoch_corret, epoch_total])
# test on local triggers
if state_key in helper.params['adversary_list']:
if helper.params['vis_trigger_split_test']:
model.trigger_agent_test_vis(vis=main.vis, epoch=epoch, acc=epoch_acc, loss=None,
eid=helper.params['environment_name'],
name=state_key + "_combine")
epoch_loss, epoch_acc, epoch_corret, epoch_total = \
test.Mytest_poison_agent_trigger(helper=helper, model=model, agent_name_key=state_key)
csv_record.poisontriggertest_result.append([state_key, state_key + "_trigger", "", epoch, epoch_loss,
epoch_acc, epoch_corret,epoch_total])
if helper.params['vis_trigger_split_test']:
model.trigger_agent_test_vis(vis=main.vis, epoch=epoch, acc=epoch_acc, loss=None,
eid=helper.params['environment_name'],
name=state_key+"_trigger")
# update the weight and bias
local_model_update_dict = dict()
for name, data in model.state_dict().items():
local_model_update_dict[name] = torch.zeros_like(data)
local_model_update_dict[name] = (data - last_params_variables[name])
last_params_variables[name] = copy.deepcopy(data)
if helper.params['aggregation_methods'] == config.AGGR_FOOLSGOLD:
epochs_local_update_list.append(client_grad)
else:
epochs_local_update_list.append(local_model_update_dict)
epochs_submit_update_dict[state_key] = epochs_local_update_list
return epochs_submit_update_dict, num_samples_dict