Vision large language models are increasingly used in privacy-sensitive domains, raising concerns about whether they expose training data. Membership inference attacks attempt to determine if a particular sample was part of the training set, and existing gray-box methods often rely on entropy statistics or generated outputs.

The new arXiv preprint presents COMPASS, a gray-box attack that calibrates noise adaptively to probe the model's behavior. This differs from earlier approaches that depend on statistical cues, offering a potentially more direct way to infer membership.

The abstract does not include full experimental details, but the method points toward noise-based probing as a promising direction for auditing privacy risks in vision LLMs.