Built an adversarial attack suite to evaluate the robustness of a pretrained ResNet-34 model under bounded perturbation attacks.
Overview
The project implements multiple attack strategies and compares how each method affects model behavior. It also evaluates transferability by testing whether adversarial examples crafted for one architecture remain effective on DenseNet-121.
Technical Focus
- FGSM, PGD, Momentum PGD, and Patch PGD implementations
- L-infinity bounded adversarial perturbation experiments
- Attack success and model degradation analysis
- Transferability testing across ResNet-34 and DenseNet-121
Repository
The implementation and evaluation notebooks are available on GitHub.