-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy pathcentralized.py
225 lines (179 loc) · 9.28 KB
/
centralized.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
# -*- coding: utf-8 -*-
"""
-----------------------------------------------
# File: centralized.py
-----------------------------------------------
"""
import numpy as np
import torch
import pandas as pd
import sys
import random
sys.path.append('../')
from utils.misc import args_parser
from utils.misc import get_data, process_isolated, set_logger, save_model
from utils.models import LSTM
from utils.cen_update import CentralUpdate, test_inference
from utils.attacks import (DataMetaPoisonSSGD, DataMetaPoisonAdam,
NoAttack, FirstOrderPoisonSSGD, UniformPoison)
from utils.defenses import DataPoisonDefense
from sklearn import metrics
import logging
import setGPU
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
if __name__ == '__main__':
args = args_parser()
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
data, _, selected_cells, mean, std, _, _ = get_data(args)
device = 'cuda' if args.gpu else 'cpu'
parameter_list = 'Centrailized-data-{:}-type-{:}-'.format(args.file, args.type)
parameter_list += '-frac-{:.2f}-epoch-{:}-batch-{:}-seed-{:}'.format(args.frac,
args.epochs,
args.batch_size,
args.seed)
log_id = parameter_list
logger = set_logger(log_id=log_id, args=args, logger_name=__name__)
train, val, test = process_isolated(args, data)
global_model = LSTM(args).to(device)
global_model.train()
global_weights = global_model.state_dict()
best_val_loss = None
val_loss = []
val_acc = []
cell_loss = []
loss_hist = []
m = max(int(args.adversary_frac * args.bs), 1)
potential_adversary_idx = random.sample(selected_cells, m)
logger.info('Potential Adversarial ID: {}'.format(' '.join(map(str, potential_adversary_idx))))
## aggregate training and testing data
'''
The structure of train[cell] is a tuple (close, period, target)
'''
train_data, test_data = [[], [], []], [[], [], []]
target_data = None
if not args.collision:
for cell in selected_cells:
if args.poison and cell in potential_adversary_idx:
if args.attack_optimizer == 'adam':
adversary = DataMetaPoisonAdam(args, train[cell], test[cell], target_data,
args.num_ensemble, attack_lr = args.attack_lr,
mask_prob = args.mask_prob)
elif args.attack_optimizer == 'ssgd':
adversary = DataMetaPoisonSSGD(args, train[cell], test[cell], target_data, args.num_ensemble,
attack_lr = args.attack_lr,
mask_prob = args.mask_prob)
elif args.attack_optimizer == 'fssgd':
adversary = FirstOrderPoisonSSGD(args, train[cell], test[cell], target_data, args.num_ensemble)
elif args.attack_optimizer == 'uniform':
adversary = UniformPoison(args, train[cell], test[cell])
else:
adversary = NoAttack(args, train[cell], test[cell], target_data, args.num_ensemble)
print('Not implemented')
# raise
train[cell] = adversary.modify_data(num_rounds=args.attack_rounds)
for idx in range(len(train_data)):
train_data[idx].append(train[cell][idx])
test_data[idx].append(test[cell][idx])
for idx in range(len(train_data)):
train_data[idx] = np.concatenate(train_data[idx], axis=0)
test_data[idx] = np.concatenate(test_data[idx], axis=0)
train_data, test_data = tuple(train_data), tuple(test_data)
else:
if args.poison:
## aggregate the poisoning data
poison_train_data, poison_test_data = [[], [], []], [[], [], []]
poison_train_idx = [0]
train_end_idx = 0
for cell in potential_adversary_idx:
for idx in range(len(poison_train_data)):
poison_train_data[idx].append(train[cell][idx].copy())
poison_test_data[idx].append(test[cell][idx].copy())
train_end_idx += train[cell][idx].shape[0]
poison_train_idx.append(train_end_idx)
for idx in range(len(poison_train_data)):
poison_train_data[idx] = np.concatenate(poison_train_data[idx], axis=0)
poison_test_data[idx] = np.concatenate(poison_test_data[idx], axis=0)
if args.attack_optimizer == 'adam':
adversary = DataMetaPoisonAdam(args, tuple(poison_train_data),
tuple(poison_test_data),
target_data, args.num_ensemble,
attack_lr = args.attack_lr,
mask_prob = args.mask_prob)
elif args.attack_optimizer == 'ssgd':
adversary = DataMetaPoisonSSGD(args, tuple(poison_train_data),
tuple(poison_test_data),
target_data, args.num_ensemble)
elif args.attack_optimizer == 'comparison':
adversary = DataMetaPoisonSSGD(args, tuple(poison_train_data),
tuple(poison_test_data),
target_data, args.num_ensemble)
else:
adversary = NoAttack(args, train[cell], test[cell], target_data, args.num_ensemble)
print('Not implemented')
# raise
poison_train_data = adversary.modify_data(num_rounds=args.attack_rounds)
for cell_idx in range(len(potential_adversary_idx)):
cell = potential_adversary_idx[cell_idx]
train[cell] = list(train[cell])
for idx in range(len(poison_train_data)):
train[cell][idx] = poison_train_data[idx][poison_train_idx[cell_idx]: \
poison_train_idx[cell_idx+1]]
train[cell] = tuple(train[cell])
for cell in selected_cells:
for idx in range(len(train_data)):
train_data[idx].append(train[cell][idx])
test_data[idx].append(test[cell][idx])
for idx in range(len(train_data)):
train_data[idx] = np.concatenate(train_data[idx], axis=0)
test_data[idx] = np.concatenate(test_data[idx], axis=0)
train_data, test_data = tuple(train_data), tuple(test_data)
if args.apply_defense is None:
pass
elif args.apply_defense == 'sphere_sani':
defense = DataPoisonDefense(args)
train_data = defense.sphere_remove_outliers(train_data, args.removal_proportion)
elif args.apply_defense == 'adj_sani':
defense = DataPoisonDefense(args)
train_data = defense.adj_remove_outliers(train_data, args.removal_proportion)
elif 'rand' in args.apply_defense:
defense = DataPoisonDefense(args)
train_data = defense.add_noise(train_data, args.sigma)
else:
raise NotImplementedError
## train the model
'''
Directly use the data to update the central model
'''
model_operation = CentralUpdate(args, train_data, test_data)
global_weights, _ = model_operation.update_weights(model=global_model, central_epochs=args.epochs)
if args.poison:
save_model(log_id, args, global_weights, selected_cells, potential_adversary_idx)
else:
save_model(log_id, args, global_weights, selected_cells)
# Test model accuracy
pred, truth = {}, {}
potential_adv_pred, potential_adv_truth = {}, {}
test_loss_list = []
test_mse_list = []
nrmse = 0.0
global_model.load_state_dict(global_weights)
for cell in selected_cells:
cell_test = test[cell]
test_loss, test_mse, test_nrmse, pred[cell], truth[cell] = test_inference(args, global_model, cell_test)
if cell in potential_adversary_idx:
# logger.info('Potential Adversarial ID: {} Target MSE: {} NRMSE: {}'.format(cell, test_mse, test_nrmse))
potential_adv_pred[cell], potential_adv_truth = pred[cell], truth[cell]
nrmse += test_nrmse
test_loss_list.append(test_loss)
test_mse_list.append(test_mse)
df_pred = pd.DataFrame.from_dict(pred)
df_truth = pd.DataFrame.from_dict(truth)
mse = metrics.mean_squared_error(df_pred.values.ravel(), df_truth.values.ravel())
mae = metrics.mean_absolute_error(df_pred.values.ravel(), df_truth.values.ravel())
nrmse = nrmse / len(selected_cells)
logger.info('Centrailized File: {:} Type: {:} MSE: {:.4f} MAE: {:.4f}, NRMSE: {:.4f}'.format(args.file, args.type, mse, mae,
nrmse))