Shape and Substance: Unmasking Privacy Leaks in On-Device AI Vision Models
Shape and Substance: Unmasking Privacy Leaks in On-Device AI Vision Models Imagine snapping a photo of your medical scan on your smartphone and asking an AI to explain it—all without sending the image to the cloud. Sounds secure, right? On-device Vision-Language Models (VLMs) like LLaVA-NeXT and Qwen2-VL make this possible, promising rock-solid privacy by keeping your data local. But a groundbreaking research paper reveals a sneaky vulnerability: attackers can peer into your photos just by watching how the AI processes them.[1] ...