-
Notifications
You must be signed in to change notification settings - Fork 1
/
run_lira_mia.py
194 lines (154 loc) · 8.55 KB
/
run_lira_mia.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
import scipy
from scipy import stats
import os
import numpy as np
import torch
import csv
from torchvision import transforms
from tqdm.auto import tqdm
from mico_competition import ChallengeDataset, load_purchase100, load_model, load_cifar10
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--challenge', type=str, required=True,choices=['cifar10', 'purchase100'])
parser.add_argument('--logit_scaling', action='store_true')
parser.add_argument('--stable_logit_scaling', action='store_true')
parser.add_argument('--global_variance', action='store_true')
parser.add_argument('--online_attack', action='store_true')
args = parser.parse_args()
CHALLENGE = args.challenge
LEN_TRAINING = 50000
LEN_CHALLENGE = 100
dataset = load_cifar10(dataset_dir=".") if CHALLENGE == 'cifar10' else load_purchase100(dataset_dir='.')
criterion = torch.nn.CrossEntropyLoss(reduction='none')
scenarios = os.listdir(CHALLENGE)
phases = ['dev', 'final', 'train']
##########################################
# store training indices for train phase #
##########################################
from collections import defaultdict
train_sets = defaultdict(dict)
for scenario in tqdm(scenarios, desc="scenario"):
root = os.path.join(CHALLENGE, scenario, 'train')
for model_folder in tqdm(sorted(os.listdir(root), key=lambda d: int(d.split('_')[1])), desc="model"):
path = os.path.join(root, model_folder)
challenge_dataset = ChallengeDataset.from_path(path, dataset=dataset, len_training=LEN_TRAINING)
challenge_points = challenge_dataset.get_challenges()
train_sets[scenario][model_folder] = challenge_dataset.member.indices + challenge_dataset.training.indices
##############################
# loading stored predictions #
##############################
predictions = defaultdict(lambda: defaultdict(dict))
trans_predictions = defaultdict(lambda: defaultdict(dict))
for scenario in tqdm(scenarios, desc="scenario"):
for phase in tqdm(phases, desc="phase"):
for i in range(100):
path = os.path.join(f'predictions_{phase}_{scenario}', f'model_{i}.npy')
predictions[phase][scenario][f'model_{i}'] = np.load(path, allow_pickle=True)[()]
#####################################
# generating membership predictions #
#####################################
for scenario in tqdm(scenarios, desc="scenario"):
for phase in tqdm(phases, desc="phase"):
root = os.path.join(CHALLENGE, scenario, phase)
for model_folder in tqdm(sorted(os.listdir(root), key=lambda d: int(d.split('_')[1])), desc="model"):
path = os.path.join(root, model_folder)
challenge_dataset = ChallengeDataset.from_path(path, dataset=dataset, len_training=LEN_TRAINING)
challenge_points = challenge_dataset.get_challenges()
challenge_dataloader = torch.utils.data.DataLoader(challenge_points, batch_size=2*LEN_CHALLENGE)
features, labels = next(iter(challenge_dataloader))
model_scores = load_model(CHALLENGE, path)(features).detach().numpy()
scores = []
means = []
predicted_scores = []
scores_in = []
means_in = []
predicted_scores_in = []
pr_out = []
for i, cp in tqdm(enumerate(challenge_points.indices), desc=f"challenge_points for {model_folder}"):
models_out = [key for key, val in train_sets[scenario].items() if cp not in val]
preds_out = np.array([predictions[phase][scenario][m][model_folder][i] for m in models_out])
############################################################
# stable logit scaling / logit scaling / no transformation #
############################################################
if args.stable_logit_scaling:
preds_out = preds_out - np.max(preds_out, axis=-1, keepdims=True)
preds_out = np.array(np.exp(preds_out), dtype=np.float64)
preds_out = preds_out / np.sum(preds_out, axis=-1, keepdims=True)
y_true = np.array([p[labels[i]] for p in preds_out])
y_wrong = np.sum(preds_out, axis=-1) - y_true
preds_out = np.log(y_true + 1e-45) - np.log(y_wrong + 1e-45)
mean_out = np.mean(preds_out)
std_out = np.std(preds_out)
model_pred = model_scores[i]
model_pred = model_pred - np.max(model_pred, keepdims=True)
model_pred = np.array(np.exp(model_pred), dtype=np.float64)
model_pred = model_pred / np.sum(model_pred, keepdims=True)
model_pred_true = model_pred[labels[i]]
model_pred_wrong = np.sum(model_pred, axis=-1) - model_pred_true
model_pred_score = np.log(model_pred_true + 1e-45) - np.log(model_pred_wrong + 1e-45)
score = model_pred_score
elif args.logit_scaling:
preds_out = scipy.special.softmax(preds_out, axis=-1)
preds_out = [p[labels[i]] for p in preds_out]
preds_out = list(map(lambda x: np.log(x / (1 - x + 10e-30)), preds_out))
mean_out = np.mean(preds_out, axis=-1)
std_out = np.std(preds_out)
score = model_scores[i]
score = scipy.special.softmax(score)[labels[i]]
score = np.log(score / (1 - score + 10e-30))
else:
preds_out = scipy.special.softmax(preds_out, axis=-1)
preds_out = [p[labels[i]] for p in preds_out]
mean_out = np.mean(preds_out, axis=-1)
std_out = np.std(preds_out)
score = model_scores[i]
score = scipy.special.softmax(score)[labels[i]]
#################
# online attack #
#################
if args.online_attack:
models_in = [key for key, val in train_sets[scenario].items() if cp in val]
preds_in = np.array([predictions[phase][scenario][m][model_folder][i] for m in models_in])
if args.stable_logit_scaling:
raise NotImplementedError('online attack with logit scaling not implemented yet')
elif args.logit_scaling:
raise NotImplementedError('online attack with logit scaling not implemented yet')
else:
if len(models_in) == 0:
mean_in = mean_out
std_in = std_out
else:
preds_in = scipy.special.softmax(preds_in, axis=-1)
preds_in = [p[labels[i]] for p in preds_in]
mean_in = np.mean(preds_in, axis=-1)
std_in = np.std(preds_in)
score_in = model_scores[i]
score_in = scipy.special.softmax(score_in)[labels[i]]
###################
# global variance #
###################
if args.global_variance:
scores.append(score)
means.append(mean_out)
predicted_scores.extend(preds_out)
if args.online_attack:
scores_in.append(score_in)
means_in.append(mean_in)
predicted_scores_in.extend(preds_in)
else:
if args.online_attack:
test_score = scipy.stats.norm.pdf(score, mean_in, std_in+1e-30) / scipy.stats.norm.pdf(score, mean_out, std_out+1e-30)
else:
test_score = scipy.stats.norm.cdf(score, mean_out, std_out+1e-30)
pr_out.append(test_score)
if args.global_variance:
if args.online_attack:
preds = scipy.stats.norm.pdf(scores_in, means_in, np.std(predicted_scores_in)+1e-30) / scipy.stats.norm.pdf(scores, means, np.std(predicted_scores)+1e-30)
else:
preds = scipy.stats.norm.cdf(scores, means, np.std(predicted_scores)+1e-30)
else:
preds = np.array(pr_out)
if not args.online_attack:
assert np.all((0 <= preds) & (preds <= 1))
with open(os.path.join(path, "prediction.csv"), "w") as f:
csv.writer(f).writerow(preds)