Feature Adversaries Remarks

[aliotopal]

While using the FeatureAdversariesTensorFlowV2 attack from ART, we noticed that the default parameters (delta=0.0, lambda=0.0, layer=-1, and max_iter=100) led to a 0% success rate. These settings do not align with the attack's original paper, which recommends using delta=10, a lambda>0, up to max_iter=500, and attacking specific internal layers such as FC7 (fully connected level 7), P5 (pooling layer 5), and C3 (convolution layer 3) instead of the undefined default layer=-1.