FRA-Attack: Frequency-Domain Method Improves Transferable Attacks on MLLMs

Researchers propose FRA-Attack, a frequency-domain regularized adversarial alignment method to improve transferable targeted attacks against closed-source multimodal large language models (MLLMs). The approach addresses two key challenges: spatial-domain feature redundancy and surrogate-specific gradient signals. By applying a high-pass DCT objective on patch features, FRA-Attack suppresses redundant global structures and focuses the loss on high-frequency bands that capture intrinsic visual focus shared across models. This enhances cross-model transferability of perturbations optimized on open-source surrogate encoders. The method is detailed in a paper on arXiv (2605.21541).