Adversarial attacks - Generative Adversarial Network (GAN) project

📄 Full PDF available here

📄 Poster available here

Code available here

Abstract

We demonstrate a purely generative approach to synthesise adversarial examples based on a U-net backed generative adversarial network. We adapt the traditional discriminator setup, adding a separate classifier in order to guide model training to produce adversarial examples without perceptual distortions or artefacts. We demonstrate the attack on the MNSIT digit dataset, contrast with existing approaches, as well as explore potential defence mechanisms.