-
Notifications
You must be signed in to change notification settings - Fork 2
/
image_train.py
466 lines (428 loc) · 19.4 KB
/
image_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
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
import copy
import os
import test
import time
import torch
import torch.nn as nn
import torch.nn.functional as F
import config
import invert_CIFAR
import invert_FashionMNIST
import invert_MNIST
import main
import utils.csv_record as csv_record
def ImageTrain(
helper, start_epoch, local_model, target_model, is_poison, agent_name_keys
):
"""
Train the local model, and return the local model and the number of samples
:param helper: the helper object
:param start_epoch: the start epoch
:param local_model: the local model
:param target_model: the target model
:param is_poison: whether the local model is poisoned
:param agent_name_keys: the agent name keys
:return: the local model, the number of samples, and the epochs_submit_update_dict
"""
epochs_submit_update_dict = dict()
num_samples_dict = dict()
current_number_of_adversaries = 0
for temp_name in agent_name_keys:
if temp_name in helper.params["adversary_list"]:
current_number_of_adversaries += 1
for model_id in range(helper.params["no_models"]):
epochs_local_update_list = []
last_local_model = dict()
client_grad = [] # only works for aggr_epoch_interval=1
for name, data in target_model.state_dict().items():
last_local_model[name] = target_model.state_dict()[name].clone()
agent_name_key = agent_name_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()
adversarial_index = -1
localmodel_poison_epochs = helper.params["poison_epochs"]
if is_poison and agent_name_key in helper.params["adversary_list"]:
for temp_index in range(0, len(helper.params["adversary_list"])):
if int(agent_name_key) == helper.params["adversary_list"][temp_index]:
adversarial_index = temp_index
localmodel_poison_epochs = helper.params[
str(temp_index) + "_poison_epochs"
]
main.logger.info(
f"poison local model {agent_name_key} index {adversarial_index} "
)
break
if len(helper.params["adversary_list"]) == 1:
adversarial_index = -1 # the global pattern
for epoch in range(
start_epoch, start_epoch + helper.params["aggr_epoch_interval"]
):
target_params_variables = dict()
for name, param in target_model.named_parameters():
target_params_variables[name] = (
last_local_model[name].clone().detach().requires_grad_(False)
)
if (
is_poison
and agent_name_key in helper.params["adversary_list"]
and (epoch in localmodel_poison_epochs)
):
main.logger.info("poison_now")
poison_lr = helper.params["poison_lr"]
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,
)
temp_local_epoch = (epoch - 1) * internal_epoch_num
for internal_epoch in range(1, internal_epoch_num + 1):
temp_local_epoch += 1
_, data_iterator = helper.train_data[agent_name_key]
poison_data_count = 0
total_loss = 0.0
correct = 0
dataset_size = 0
dis2global_list = []
for batch_id, batch in enumerate(data_iterator):
data, targets, poison_num = helper.get_poison_batch(
batch, adversarial_index=adversarial_index, evaluation=False
)
poison_optimizer.zero_grad()
dataset_size += len(data)
poison_data_count += poison_num
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()
)
if step_lr:
scheduler.step()
main.logger.info(f"Current lr: {scheduler.get_lr()}")
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,
agent_name_key,
internal_epoch,
total_l,
correct,
dataset_size,
acc,
poison_data_count,
)
)
csv_record.train_result.append(
[
agent_name_key,
temp_local_epoch,
epoch,
internal_epoch,
total_l.item(),
acc,
correct,
dataset_size,
]
)
num_samples_dict[agent_name_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)}"
)
if not helper.params["baseline"]:
main.logger.info(f"will scale.")
(
epoch_loss,
epoch_acc,
epoch_was_corret,
epoch_corret,
epoch_total,
) = test.Mytest(
helper=helper,
epoch=epoch,
model=model,
is_poison=False,
agent_name_key=agent_name_key,
)
csv_record.test_result.append(
[
agent_name_key,
epoch,
epoch_loss,
epoch_acc,
epoch_was_corret,
epoch_corret,
epoch_total,
]
)
(
epoch_loss,
epoch_acc,
epoch_was_corret,
epoch_corret,
epoch_total,
) = test.Mytest_poison(
helper=helper,
epoch=epoch,
model=model,
is_poison=True,
agent_name_key=agent_name_key,
)
csv_record.posiontest_result.append(
[
agent_name_key,
epoch,
epoch_loss,
epoch_acc,
epoch_was_corret,
epoch_corret,
epoch_total,
]
)
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_local_model[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}"
)
csv_record.scale_temp_one_row.append(epoch)
csv_record.scale_temp_one_row.append(round(distance, 4))
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:
if helper.params["invert"]:
_, data_iterator = helper.train_data[agent_name_key]
main.logger.info(
"___Train {}, epoch {:3d}, local model {}".format(
model.name, epoch, agent_name_key
)
)
if helper.params["type"] == config.TYPE_CIFAR:
model, dataset_size = invert_CIFAR.trigger_fast_train(
helper, model, data_iterator, start_epoch, agent_name_key
)
elif helper.params["type"] == config.TYPE_MNIST:
model, dataset_size = invert_MNIST.trigger_fast_train(
helper, model, data_iterator, start_epoch, agent_name_key
)
elif helper.params["type"] == config.TYPE_FASHION_MNIST:
model, dataset_size = invert_FashionMNIST.trigger_fast_train(
helper, model, data_iterator, start_epoch, agent_name_key
)
main.logger.info(
f"==== after train, get the dataset_size: {dataset_size}"
)
num_samples_dict[agent_name_key] = dataset_size
else: # normal training
temp_local_epoch = (epoch - 1) * helper.params["internal_epochs"]
for internal_epoch in range(
1, helper.params["internal_epochs"] + 1
):
temp_local_epoch += 1
_, data_iterator = helper.train_data[agent_name_key]
total_loss = 0.0
correct = 0
dataset_size = 0
dis2global_list = []
for batch_id, batch in enumerate(data_iterator):
optimizer.zero_grad()
data, targets = helper.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,
agent_name_key,
internal_epoch,
total_l,
correct,
dataset_size,
acc,
)
)
csv_record.train_result.append(
[
agent_name_key,
temp_local_epoch,
epoch,
internal_epoch,
total_l.item(),
acc,
correct,
dataset_size,
]
)
num_samples_dict[agent_name_key] = dataset_size
# test local model after internal epoch finishing
(
epoch_loss,
epoch_acc,
epoch_was_corret,
epoch_corret,
epoch_total,
) = test.Mytest(
helper=helper,
epoch=epoch,
model=model,
is_poison=False,
agent_name_key=agent_name_key,
)
csv_record.test_result.append(
[
agent_name_key,
epoch,
epoch_loss,
epoch_acc,
epoch_was_corret,
epoch_corret,
epoch_total,
]
)
if is_poison:
if agent_name_key in helper.params["adversary_list"] and (
epoch in localmodel_poison_epochs
):
(
epoch_loss,
epoch_acc,
epoch_was_corret,
epoch_corret,
epoch_total,
) = test.Mytest_poison(
helper=helper,
epoch=epoch,
model=model,
is_poison=True,
agent_name_key=agent_name_key,
)
csv_record.posiontest_result.append(
[
agent_name_key,
epoch,
epoch_loss,
epoch_acc,
epoch_was_corret,
epoch_corret,
epoch_total,
]
)
# test on local triggers
if agent_name_key in helper.params["adversary_list"]:
(
epoch_loss,
epoch_acc,
epoch_was_corret,
epoch_corret,
epoch_total,
) = test.Mytest_poison_agent_trigger(
helper=helper, model=model, agent_name_key=agent_name_key
)
csv_record.poisontriggertest_result.append(
[
agent_name_key,
str(agent_name_key) + "_trigger",
"",
epoch,
epoch_loss,
epoch_acc,
epoch_was_corret,
epoch_corret,
epoch_total,
]
)
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_local_model[name]
last_local_model[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[agent_name_key] = epochs_local_update_list
return epochs_submit_update_dict, num_samples_dict