This repository contains my work for my Master's Thesis titled ARCANE: Adversarial Robustness using Class-conditionAl geNerative modEls

ARCANE is a novel framework whose aim is to provide adversarial robustness to classifier models using class-conditional generative models. By projecting a given potentially adversarial sample onto the distribution learned through a generative model, ARCANE is able to purify a given sample to produce clean class labels. Moreover, a small classifier is able to detect whether a given sample contains adversarial noise with a set of 6 features that are extracted from the sample at inference time.