Featurized Bidirectional GAN: Adversarial Defense via Adversarially Learned Semantic Inference

no code implementations • ICLR 2019 • Ruying Bao, Sihang Liang, Qingcan Wang

In this paper, we propose a defense method, Featurized Bidirectional Generative Adversarial Networks (FBGAN), to extract the semantic features of the input and filter the non-semantic perturbation.

Adversarial Defense