#artificial intelligence
Feb 2026
Robustness of Vision Language Models Against Split-Image Harmful Input Attacks
This work introduces novel split-image visual jailbreak attacks (SIVA) that exploit a new vulnerability in VLM safety alignment and introduces a novel adversarial knowledge distillation algorithm to substantially improve cross-model transferability.
Md. Rafi Ur Rashid, Sadik Hossain Shanto, Vishnu Asutosh Dasu et al.
· arXiv.org · 1 citation