FCAT: Federated causal adversarial training

Abstract

Causal inference has been proven to be a crucial technique for improving the efficacy and explainability of adversarial training (AT). However, its applicability in the decentralized adversarial training paradigm has not been fully explored. Where one potential challenge is to apply the causal inference in the settings of non-independent and identically distributed (Non-IID) federated learning. In particular, the imbalanced data distributions among various clients will unavoidably hinder the efficacy and adaptability of causal inference. To address this issue, this paper proposes a novel yet practical method dubbed Federated Causal Adversarial Training (FCAT), which seeks to improve causal models via calibrated correction information. Additionally, we introduce a lightweight slack aggregation method aimed at addressing client model disparities and minimizing the communication overhead in each iteration. Extensive experimental results demonstrate that FCAT significantly improves the efficacy of causal models in federated adversarial training, and remarkably outperforms the current state-of-the-art (SOTA) competitors on multiple widely-adopted benchmarks.