Exploring Diversified Adversarial Robustness in Neural Networks via Robust Mode Connectivity

This paper proposes a new method called robust mode connectivity (RMC) to enhance the adversarial robustness of neural networks (NNs) by exploring a wider range of parameter space. While adversarial training methods have shown promising results in enhancing the robustness of NNs against perturbations, they are limited by considering only a single type of perturbation during training and having limited search capability. RMC aims to address this limitation by considering multiple ℓ
p
norm perturbations (p = 1, 2, ∞) and building on the concept of mode connectivity to identify a path of NNs with high robustness against different types of perturbations. The proposed method employs a multi steepest descent (MSD) algorithm to explore the parameter space and achieve diversified adversarial robustness. Experimental results on various datasets and architectures demonstrate the effectiveness of RMC.