kingkenche/Art_Ass_5_q2

Assignment 5 – Adversarial Attacks & Detection (Q2)

Overview

• ResNet-18 trained from scratch on CIFAR-10 (≥ 72% clean accuracy).
• FGSM attack implemented from scratch and via IBM ART.
• ResNet-34 binary detectors trained on PGD and BIM adversarial examples.

Checkpoints

File	Description
checkpoints/resnet18_cifar10_best.pth	Best ResNet-18 classifier
checkpoints/detector_pgd.pth	ResNet-34 PGD detector
checkpoints/detector_bim.pth	ResNet-34 BIM detector

Results

See results/ folder and the WandB project linked in the GitHub README.