Researchers have identified a new vulnerability in Vision-Language Models (VLMs) where safety alignment is not robust to harmful content split across multiple image fragments. While VLMs generalize well to split images during pretraining, their safety mechanisms, often trained on holistic images, fail to detect combined harmful semantics. The study introduces novel split-image visual jailbreak attacks (SIVA) that progressively evolve from simple splitting to adaptive white-box and black-box transfer attacks, utilizing adversarial knowledge distillation (Adv-KD) to enhance cross-model transferability. Evaluations showed these attacks achieved significantly higher success rates than existing methods on state-of-the-art VLMs. AI
IMPACT Highlights a critical gap in VLM safety alignment, potentially requiring new training methodologies to address split-image attacks.
RANK_REASON Academic paper detailing a new vulnerability and attack method for VLMs. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
