Using vision transformer-based GANs against Vision Transformers

Andrei Florin Pămînt, Sergiu Adrian Darabant

Keywords: ViTs, Adversarial, Attack, ViTGAN, A-ViTGAN

TL;DR: We introduce a single encoder and a three-encoder Transformer based GAN that creates a perturbation with a successful attack rate higher than state of the art methods.

Abstract: Vision transformers have become one of the best architectures for image classification tasks. In this paper, we introduce a novel method for creating adversarial attacks in a black box environment without using surrogate models. Specifically, we introduce a single encoder and a three-encoder Transformer based GAN that creates a perturbation with a successful attack rate higher than state of the art methods.