Biased Multi-Domain Adversarial Training

Saehyung Lee, Hyungyu Lee, Sanghyuk Chun, Sungroh Yoon

Keywords: adversarial training, adversarial robustness

Abstract: Several recent studies have shown that the use of extra in-distribution data can lead to a high level of adversarial robustness. However, there is no guarantee that it will always be possible to obtain sufficient extra data for a selected dataset. In this paper, we propose a biased multi-domain adversarial training (BiaMAT) method that induces training data amplification using freely available auxiliary datasets. The proposed method can achieve increased adversarial robustness on a primary dataset by leveraging auxiliary datasets via multi-domain learning. Specifically, data amplification on both robust and non-robust features can be accomplished through the application of BiaMAT as demonstrated through an additional analysis based on shuffle testing. Our experimental results indicate that BiaMAT can effectively utilize the robust and non-robust features present in various auxiliary datasets. Moreover, we demonstrate that while existing methods are vulnerable to negative transfer due to the distributional discrepancy between auxiliary and primary data, the proposed method enables neural networks to flexibly leverage diverse image datasets for adversarial training by successfully handling the domain discrepancy through the application of a confidence-based selection strategy.

One-sentence Summary: we propose a biased multi-domain adversarial training (BiaMAT) method that can achieve increased adversarial robustness on a primary dataset by leveraging auxiliary datasets via multi-domain learning.