Steganografiya quyidagi sohalarda qo'llaniladi, lekin ular bilan cheklanmaydi

test_size=0.1,
random_state=42)

# Train a watermarked model
print('Training watermarked model')
number_labels = len(np.unique([floor(k) for k in y_train]))
args = TrainingWMArgs(
nbr_classes=number_labels,
trigger_size=trigger_size,
trigger_technique='dawn',
metric=metric)

wm_model = Trainer(clone(base_model), args=args)
ownership = wm_model.fit(X_train, y_train)
WM_X = ownership['inputs']
number_labels = len(np.unique([floor(k) for k in y_train]))

# Train a non-watermarked model
print('Training non-watermarked model')
clean_model = clone(base_model)
clean_model.fit(X_train, y_train)

# Verification for non-stolen
print('Clean model not detected as stolen...', end=' ')
if metric == 'accuracy':
verification = verify(
ownership['labels'],
clean_model.predict(WM_X),
number_labels=number_labels,
metric=metric)

else:
verification = verify(
ownership['labels'],
clean_model.predict(WM_X),
number_labels=ownership['selected_q'],
bounds=(min(y), max(y)),
metric=metric)
assert not verification['is_stolen']
print('Done!')

# Verification for stolen
print('Stolen watermarked model detected as stolen...', end=' ')
if metric == 'accuracy':
verification = verify(
ownership['labels'],
wm_model.predict(WM_X),
number_labels=number_labels,
metric=metric)
else:
verification = verify(
ownership['labels'],
wm_model.predict(WM_X),
number_labels=ownership['selected_q'],