The world is being quietly rearranged by people who write very long documents.


The title they went with Shape and Substance: Dual-Layer Side-Channel Attacks on Local Vision-Language Models Noisy translates that to

Local AI models leak private image details through processing speed


Researchers found that when AI image models run locally on devices (supposed to be private), the way they break down images into pieces creates detectable patterns in how fast they process—revealing what kind of image you're looking at. An attacker without special privileges can watch how long the processing takes and infer whether you're viewing a medical X-ray, a text document, or something else entirely, defeating the privacy promise of on-device execution.
Companies are moving AI image models to run on your device instead of their servers to claim privacy, but this research shows that claim is false—the way these models work creates unavoidable leaks that let attackers see what you're processing even without breaking into the model itself. This is a structural vulnerability in a design pattern that's becoming standard, not a bug in one product.

If you insist
Read the original →