Adversarial Training (ECE-653)

Term Project - Testing impacts of Adversarial Training on models' susceptibility to Membership Inference Attacks

Objectives

1. Use 2 datasets (MNIST, CIFAR-10) to train 2 models (LeNet-5, ResNet)
2. Gather metrics on the models' performance on clean data, Adversarial samples (FGSM, PGD), and against Membership Inference attack
3. Apply Adversarial Training Defenses (Adversarial Training, Auxiliary Binary Detection, K+1 Auxiliary Model, Double-Boundary Adversarial Training) to models
4. Gather metrics on models' performance on clean data, Adversarial Samples (FGSM, PGD), and against Membership Inference attacks after Adversarial Training

This repo is a copy of my school-related private repo. This project was done in collaboration with Ayra Islam.