Browse Teach Curate

Image-to-Video Surrogate Attacks for Transferable Adversarial Examples on Video Multimodal LLMs

A black-box attack that crafts adversarial videos able to fool unseen video multimodal language models. It uses an image-based multimodal model as a surrogate, perturbing frames by projected gradient descent to disrupt the latent video representation while fusing textual and spatiotemporal cues, and propagates perturbations across frames to survive unknown frame-sampling schemes.

KITT Text

Transferability of Adversarial Attacks in Video-based MLLMs: A Cross-modal Image-to-Video Approach

Video-based multimodal large language models (V-MLLMs) answer questions about videos but can be fooled by adversarial videos -- clips altered with small, optimized pixel perturbations that change the…