Rethinking Robust Contrastive Learning from the Adversarial Perspective
This work addresses the problem of improving robustness in deep learning networks for researchers and practitioners, offering insights into representation convergence, but it appears incremental as it builds on existing adversarial training methods.
The paper investigates how adversarial training affects contrastive learning and supervised learning, finding that it reduces disparities between adversarial and clean representations and improves robustness by increasing their similarity, particularly in later network layers.
To advance the understanding of robust deep learning, we delve into the effects of adversarial training on self-supervised and supervised contrastive learning alongside supervised learning. Our analysis uncovers significant disparities between adversarial and clean representations in standard-trained networks across various learning algorithms. Remarkably, adversarial training mitigates these disparities and fosters the convergence of representations toward a universal set, regardless of the learning scheme used. Additionally, increasing the similarity between adversarial and clean representations, particularly near the end of the network, enhances network robustness. These findings offer valuable insights for designing and training effective and robust deep learning networks. Our code is released at \textcolor{magenta}{\url{https://github.com/softsys4ai/CL-Robustness}}.