Tinsae/Finetuning
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Black box

Browse Source
This commit is contained in:
Tinsae
2026-07-08 13:55:02 +02:00
parent 945d0298d0
commit 90e33f074d
2 changed files with 20 additions and 74 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
eval/UnlearningAttack.py
View File

@@ -18,9 +18,9 @@ class UnlearningAttack:
self.arch = arch
self.class_size = class_size
self.hook = None
self.model = None
self._hook_features = []
#self.hook = None
#self.model = None
#self._hook_features = []
self.criterion = nn.CrossEntropyLoss(reduction='none')
self.collecting = False
@@ -46,7 +46,7 @@ class UnlearningAttack:
def calculate_a_dist(self, latent1, latent2):
"""Calculates formal A-Distance: 2 * (1 - 2 * epsilon)."""
combined = np.vstack([latent1, latent2])
'''combined = np.vstack([latent1, latent2])
mean = np.mean(combined, axis=0)
std = np.std(combined, axis=0) + 1e-8
l1 = (latent1 - mean) / std
@@ -63,13 +63,16 @@ class UnlearningAttack:
split = int(len(X) * 0.7)
clf = LogisticRegression(solver='liblinear').fit(X[:split], y[:split])
epsilon = 1.0 - accuracy_score(y[split:], clf.predict(X[split:]))
epsilon = 1.0 - accuracy_score(y[split:], clf.predict(X[split:]))'''
accuracy_score = self._comput_adversarial_accuracy(latent1, latent2, axis=0)
epsilon = 1 - accuracy_score
return 2.0 * np.abs(0.5 - epsilon)
def _hook_fn(self, module, input, output):
'''def _hook_fn(self, module, input, output):
if not self.collecting:
return
flattened_embeddings = torch.flatten(output, 1)
@@ -99,9 +102,11 @@ class UnlearningAttack:
if hasattr(self, 'hook') and self.hook:
self.hook.remove()
self.hook = None
'''
def _extract_attack_features(self, target_model, loader, device, target_class):
'''target_model.eval()
'''
# cnahgng to black box.
target_model.eval()
all_probs = []
all_entropies = []
all_losses = []
@@ -223,11 +228,11 @@ class UnlearningAttack:
# Note: Latent distance is removed as it's not a black-box metric
return mia_accuracy, 0.0
def _comput_adversarial_accuracy(self, filtered, naive, axis=-1):
def _comput_adversarial_accuracy(self, filtered, naive, axis=0):
# Z-Score Normalisation
filtered = (filtered - np.mean(filtered, axis=-1, keepdims=True)) / (np.std(filtered, axis=-1, keepdims=True) + 1e-8)
naive = (naive - np.mean(naive, axis=-1, keepdims=True)) / (np.std(naive, axis=-1, keepdims=True) + 1e-8)
filtered = (filtered - np.mean(filtered, axis = axis, keepdims=True)) / (np.std(filtered, axis = axis, keepdims = True) + 1e-8)
naive = (naive - np.mean(naive, axis = axis, keepdims = True)) / (np.std(naive, axis = axis, keepdims = True) + 1e-8)
# shuffle indices
num_images = len(filtered)
@@ -282,7 +287,7 @@ class UnlearningAttack:
naive = torch.cat(naive_logits, dim=0).cpu().numpy()
# evaluate similarity of outputs
lookalike_accuracy = self._comput_adversarial_accuracy(filtered=filtered, naive=naive)
lookalike_accuracy = self._comput_adversarial_accuracy(filtered=filtered, naive=naive, axis = -1)
# so that the metric is between 0 and 1.
return 2.0 * np.abs(lookalike_accuracy - 0.5)
@@ -296,7 +301,7 @@ class UnlearningAttack:
with open(current_log_file, "w") as f:
f.write("target_class, parameter_mia_accuracy, latent_distance_tell, lookalike_accuracy, A-Dist, JS-Dist\n")
self.register_model_hook(unlearned_instance.model)
#self.register_model_hook(unlearned_instance.model)
# 1. Parameter-Space MIA and Latent Distance
parameter_mia_acc, latent_dist = self.run_parameter_space_mia(