← Image classification · Repository
This folder contains standalone TensorFlow attack/interpretability scripts:
adv_sal.py, black_box.py, fgsm.py, and illm.py. They use model
gradients or score queries to construct adversarial perturbations or saliency
signals; they are not a shared benchmark suite.
| Script | Technique represented in source | Core operation |
|---|---|---|
adv_sal.py |
Adversarial saliency map | Differentiate selected class scores with respect to input features |
fgsm.py |
Fast Gradient Sign Method | Add an epsilon-scaled sign of the input gradient |
illm.py |
Iterative least-likely method | Iteratively optimise an input toward a least-likely target |
black_box.py |
Black-box attack experiments | Query-driven perturbation/search code with TensorFlow variables |
For an input (x), loss (L), and perturbation budget (\epsilon), FGSM is
the familiar tensor update (x' = x + \epsilon,\mathrm{sign}(\nabla_x L)).
The scripts use tf.GradientTape where gradients are available. Their
computational bottleneck is repeated forward/backward evaluation, so batches
and GPU placement matter more than Python control flow.
No retained accuracy/attack-success table is present; use the scripts as implementation studies and record any new evaluation externally.