Mordechai Guri

This paper introduces a dual-layer side-channel attack framework that exploits the variable workload introduced by dynamic image preprocessing in local Vision-Language Models (VLMs) to infer sensitive input content.