EMO-GAT: An Evolutionary Multi-Objective Method for Adversarial Training

Adversarial examples are an effective method that can be leveraged to attack machine learning systems. The primary defensive method, adversarial training, is a simplistic approach when applied without optimization. In this paper we propose a novel method, evolutionary multi-objective optimization genetic adversarial training (EMO-GAT), that enables the development of a more robust model by performing evolutionary adversarial training using multiple types of adversarial examples. We demonstrate that EMO-GAT outperforms both traditional adversarial training and genetic adversarial training (GAT).