def main():
if not os.path.exists(args.output):
os.makedirs(args.output)
model, img_shape = getmodel(args.model)
attacker = attack.FGSM(model=model,
goal=args.goal,
distance_metric=args.distance,
eps=args.eps)
path = os.path.join('data', 'cfp-{}x{}'.format(img_shape[0], img_shape[1]))
run_black(path, attacker, model, args)
def main():
if not os.path.exists(args.output):
os.makedirs(args.output)
model, img_shape = getmodel(args.model)
attacker = attack.Evolutionary(
model=model,
goal=args.goal,
distance_metric=args.distance,
max_queries=args.iters,
threshold=threshold[args.model]['cos'],
)
path = os.path.join('data', 'cfp-{}x{}'.format(img_shape[0], img_shape[1]))
Attacker = lambda xs, ys, pairs: attacker.batch_attack(xs, ys, pairs=pairs)
run_white(path, Attacker, model, args, threshold[args.model]['cos'])
def main():
if not os.path.exists(args.output):
os.makedirs(args.output)
model, img_shape = getmodel(args.model)
attacker = attack.LGC(model=model,
goal=args.goal,
distance_metric=args.distance,
eps=args.eps,
mu=args.mu,
num_samples=args.num_samples,
iters=args.iters)
path = os.path.join('data', 'ytf-{}x{}'.format(img_shape[0], img_shape[1]))
run_black(path, attacker, model, args)
def main():
if not os.path.exists(args.output):
os.makedirs(args.output)
model, img_shape = getmodel(args.model)
attacker = attack.CW(model=model,
goal=args.goal,
distance_metric=args.distance,
iteration=args.iters,
threshold=threshold[args.model]['cos'],
search_steps=args.steps,
binsearch_steps=args.bin_steps,
confidence=args.confidence,
learning_rate=args.lr,
c=args.c)
path = os.path.join('data', 'lfw-{}x{}'.format(img_shape[0], img_shape[1]))
Attacker = lambda xs, ys, pairs: attacker.batch_attack(xs, ys, pairs=pairs)
run_white(path, Attacker, model, args, threshold[args.model]['cos'])
def main():
model, img_shape = getmodel(args.model)
config = {
'eps': args.eps,
'method': attack.FGSM,
'goal': args.goal,
'distance_metric': args.distance,
'threshold': threshold_ytf[args.model]['cos'],
'steps': args.steps,
'bin_steps': args.bin_steps,
'model': model,
}
path = os.path.join('data', 'ytf-{}x{}'.format(img_shape[0], img_shape[1]))
Attacker = lambda xs, ys, pairs: binsearch_basic(
xs=xs, ys=ys, pairs=pairs, **config)
run_white(path, Attacker, model, args, threshold_ytf[args.model]['cos'])
def main():
if not os.path.exists(args.output):
os.makedirs(args.output)
model, img_shape = getmodel(args.model)
config = {
'eps': args.eps,
'method': attack.BIM,
'goal': args.goal,
'distance_metric': args.distance,
'threshold': threshold[args.model]['cos'],
'steps': args.steps,
'bin_steps': args.bin_steps,
'model': model,
'iters': args.iters,
}
path = os.path.join('data', 'lfw-{}x{}'.format(img_shape[0], img_shape[1]))
Attacker = lambda xs, ys, pairs: binsearch_alpha(xs=xs, ys=ys, pairs=pairs, **config)
run_white(path, Attacker, model, args, threshold[args.model]['cos'])
type=str,
default='impersonate',
choices=['dodging', 'impersonate'])
parser.add_argument('--distance',
help='l2 or linf',
type=str,
default='linf',
choices=['linf', 'l2'])
parser.add_argument('--anno',
help='output dir',
type=str,
default='output/expdemo/annotation.txt')
parser.add_argument('--log', help='log file', type=str, default='log.txt')
args = parser.parse_args()
model, img_shape = getmodel(args.model)
outputs = []
with open(args.anno, 'r') as f:
for line in tqdm(f.readlines(), total=2500):
adv_img_path, src_img_path, tar_img_path = line.strip().split(' ')
adv_img = np.load(adv_img_path)
resized_adv_img = cv2.resize(adv_img, (img_shape[1], img_shape[0]))
resized_adv_img = torch.Tensor(
resized_adv_img.transpose(2, 0, 1)[None, :]).cuda()
adv_feat = model.forward(resized_adv_img)
src_img = imread(src_img_path).astype(np.float32)
tar_img = imread(tar_img_path).astype(np.float32)
resized_tar_img = cv2.resize(tar_img, (img_shape[1], img_shape[0]))
