AI RESEARCH
V-Attack: Targeting Disentangled Value Features for Controllable Adversarial Attacks on LVLMs
arXiv CS.CV
•
ArXi:2511.20223v2 Announce Type: replace Adversarial attacks have evolved from simply disrupting predictions on conventional task-specific models to the complex goal of manipulating image semantics on Large Vision-Language Models (LVLMs). However, existing methods struggle with controllability and fail to precisely manipulate the semantics of specific concepts in the image.