def attack_statistics(args, model, x_test, distances, perturbed, or_acc,
m_train):
attacks = list(distances.keys())
SR = {}
ptb = {}
ptbr = {}
confidence = {}
for a in attacks:
ind = ind_perturbed(distances[a])
SR[a] = np.round(
100 * np.sum(
np.logical_and(np.greater(distances[a], 0),
~np.isinf(distances[a]))) / or_acc, 2)
if len(ind) > 0:
x_adv = perturbed[a][ind]
confidence[a] = np.max(model.predict(x_test - m_train,
batch_size=args.batch_size),
axis=-1)
ptb[a] = distance(x_test[ind], x_adv, 'ptb')
ptbr[a] = distance(x_test[ind], x_adv, 'ptbr')
return SR, confidence, ptb, ptbr
def evaluate(x_test, distances, perturbed, or_acc):
ind = ind_perturbed(distances)
x_adv = perturbed[ind]
l2_dist = np.zeros(len(distances))
l2_dist[ind] = distance(x_test[ind], x_adv, 2)
ind_non_pert = np.where(distances == np.inf)[0]
l2_dist[ind_non_pert] = np.inf
unique_dist, counts = np.unique(l2_dist, return_counts=True)
sorted_dist = np.zeros_like(l2_dist)
acc = or_acc * np.ones_like(sorted_dist)
tot_c = 0
for i in range(len(unique_dist)):
sorted_dist[tot_c:tot_c + counts[i]] = unique_dist[i]
if unique_dist[i] == np.inf:
acc[tot_c: tot_c + counts[i]] = acc[max(tot_c - 1, 0)]
elif unique_dist[i] > 0:
acc[tot_c: tot_c + counts[i]] = acc[max(tot_c - 1, 0)] - (np.arange(counts[i]) + 1)
tot_c += counts[i]
return sorted_dist, acc/len(x_test)
def run_attacks(args, filepath, x_test, y_test, y_test_onehot, attacks, m_train, save_all=True):
base_model_name = args.model_name
if args.extension is not None:
base_model_name = re.sub('_' + args.extension, '', base_model_name)
if not args.pretrained_weights:
model = load_model(filepath['models'] + filepath['dataset'] + args.model_name + '.h5')
else:
model = select_model(x_test.shape[1:], base_model_name, SGD(0.001, momentum=0.9, nesterov=True), args.weight_decay)
print(f"Loading pretrained weights for model: {base_model_name}")
model.load_weights(filepath['trained_weights'] + filepath['dataset'] + args.model_name + '.h5')
if args.pixel_mean:
fb_model = foolbox.models.TensorFlowModel.from_keras(model=model, bounds=(0., 1.), preprocessing={'mean': m_train})
else:
m_train = 0
fb_model = foolbox.models.TensorFlowModel.from_keras(model=model, bounds=(0., 1.))
print("Model: ", args.model_name)
test_loss, test_acc = model.evaluate(x_test - m_train, y_test_onehot, batch_size=args.batch_size, verbose=0)
y_pred = np.argmax(model.predict(x_test - m_train, batch_size=args.batch_size), axis=-1)
or_acc = np.sum(y_test == y_pred)
print(f"Test loss: {test_loss}, test acc: {test_acc}")
# print(f"Or acc: ", or_acc)
# For extracting statistics. Initialising with empty lists
if os.path.exists(filepath['output'] + filepath['dataset'] + args.model_name + '.npz'):
print(f"Resuming attacks for model: {args.model_name}")
npzfile = np.load(filepath['output'] + filepath['dataset'] + args.model_name + '.npz', allow_pickle=True)
labels = npzfile['arr_0'].item()
perturbed = npzfile['arr_1'].item()
distances = npzfile['arr_2'].item()
l2_distances = npzfile['arr_3'].item()
sorted_distances = npzfile['arr_4'].item()
adv_acc = npzfile['arr_5'].item()
dict_loss = npzfile['arr_9'].item()
dict_acc = npzfile['arr_10'].item()
else:
print(f"Starting attack from scratch for model: {args.model_name}")
labels = {}
perturbed = {}
distances = {}
l2_distances = {}
sorted_distances = {}
adv_acc = {}
dict_loss = {}
dict_acc = {}
for r in range(args.repetitions):
batch_ind = batch_generator(len(y_test), batch_size=args.adversarial_batch_size, shuffle=args.shuffle)
# print("B: ", batch_ind)
for a in attacks:
label = []
dist = []
pert = []
fb_attack = foolbox_attack(a, fb_model)
for b in batch_ind:
if a == 'ShiftsAttack':
adversarials = fb_attack(x_test[b], y_test[b], unpack=False, do_rotations=False)
else:
adversarials = fb_attack(x_test[b], y_test[b], unpack=False)
label.extend([ad.adversarial_class for ad in adversarials])
dist.extend([ad.distance.value for ad in adversarials])
pert.extend([ad.perturbed for ad in adversarials])
# print("Label shape: ", len(label))
label = np.array(label)
dist = np.array(dist)
pert = np.array(pert)
# print(f"label: {label.shape}, dist: {dist.shape}, pert: {pert.shape}")
if args.shuffle:
unshuffle_ind = unshuffle_index(batch_ind)
label = label[unshuffle_ind]
dist = dist[unshuffle_ind]
pert = pert[unshuffle_ind]
ind_pert = ind_perturbed(dist)
ind_not_inf = ind_not_infinite(dist)
dist = dist.astype(np.float32)
if label.dtype == 'object':
label = convert_object_array(label, in_shape=len(x_test), ind_pert=ind_not_inf, dtype=np.int8)
pert = convert_object_array(pert, in_shape=x_test.shape[1:], ind_pert=ind_not_inf, dtype=np.float32)
x_adv = np.array(x_test, copy=True)
if len(ind_pert) > 0:
x_adv[ind_pert] = pert[ind_pert]
l2_dist = distance(x_test, x_adv, 2)
if a not in labels.keys():
labels[a] = label
distances[a] = dist
perturbed[a] = pert
l2_distances[a] = l2_dist
else:
ind = np.where(np.logical_and(l2_dist > 0, np.logical_or(l2_dist < l2_distances[a], l2_distances[a] == 0)))[0]
if np.sum(ind) > 0:
labels[a][ind] = label[ind]
distances[a][ind] = dist[ind]
# print("Pert: ", pert[ind].shape)
# print("PP: ", perturbed[a][ind].shape)
perturbed[a][ind] = pert[ind]
l2_distances[a][ind] = l2_dist[ind]
if r == args.repetitions-1:
sorted_dist, acc = evaluate(x_test, distances[a], perturbed[a], or_acc)
sorted_distances[a] = sorted_dist
adv_acc[a] = acc
if args.ablation:
dict_loss, dict_acc = ablation(args, model, x_test-m_train, y_test_onehot, distances, perturbed, m_train)
sorted_distances, adv_acc = to_nparray(sorted_distances), to_nparray(adv_acc)
SR, confidence, ptb, ptbr = attack_statistics(args, model, x_test, distances, perturbed, or_acc, m_train)
if save_all:
np.savez(filepath['output'] + filepath['dataset'] + args.model_name,
labels, perturbed, distances, l2_distances, sorted_distances, adv_acc,
SR, ptb, ptbr, dict_loss, dict_acc)
return labels, perturbed, distances, l2_distances, sorted_distances, adv_acc, SR, ptb, ptbr, dict_loss, dict_acc