Cycle monotonicity of adversarial attacks for optimal domain adaptation

Abstract

We reveal an intriguing connection between adversarial attacks and cycle monotone maps, also known as optimal transport maps. Based on this finding, we developed a novel method named source fiction for semi-supervised optimal transport-based domain adaptation. In our algorithm, instead of mapping from target to the source domain, optimal transport maps target samples to the set of adversarial examples. The trick is that these adversarial examples are labeled target samples perturbed to look like source samples for the source domain classifier. Due to the cycle monotonicity of adversarial attacks, optimal transport can naturally approximate this transformation. We conduct experiments on various datasets and show that our method can notably improve the performance of optimal transport methods in semi-supervised domain adaptation.

Tasks

Reproductions