DistractMIA: Black-Box Membership Inference Attack on Vision-Language Models

A new approach called DistractMIA has been introduced by researchers as a black-box membership inference attack targeting vision-language models (VLMs), relying solely on generated text responses. This technique differs from earlier strategies that depended on probability-level indicators or mask-based outputs. Instead, DistractMIA integrates a semantic distractor into the original image and observes the variations in the model's responses. The rationale behind this method is the belief that member samples are more closely tied to the semantics of the original image. The research paper can be found on arXiv with the identifier 2605.12574.