Adversarial vulnerability in deep neural networks (DNNs) has been studied from various perspectives including decision-boundary geometry, feature robustness, input-output Jacobians, and the instability of inverse problems. This paper focuses on the spectral structure of intermediate linear transformations in modern DNNs, which is an unexplored mechanism of adversarial vulnerability. We specifically investigate transformer-based vision-language models (VLMs), whose linear layers allow for interpretable spectral decompositions, making the understanding of their robustness increasingly important. We propose a white-box spectral-subspace-guided attack (SSGRA) that aligns intermediate representations with the subspace spanned by the bottom right singular vectors. Our experiments show that SSGRA enhances attack effectiveness compared to existing baselines. Furthermore, SSGRA offers a spectral interpretation of adversarial vulnerability in VLMs, providing insights for improving their robustness.
Blogger's Review: This paper provides a fresh perspective and methodology for adversarial attacks on vision-language models by exploring intermediate spectral subspaces. As DNN applications grow, understanding their vulnerability mechanisms is crucial for enhancing model safety and robustness. The introduction of the SSGRA method showcases the potential for integration between theory and practice, warranting further exploration and application.