Amplifying Membership Signal Through Iterative Regeneration

The tendency of large generative models to memorize training data has established sample verification as a critical necessity for privacy auditing and copyright enforcement. Current membership inference attacks (MIAs) often rely on "one-shot" generations, which yield weak signals and limited sensitivity across different modalities. Inspired by Model Autophagy Disorder (MAD), we introduce MADreMIA, a model-agnostic add-on framework that enhances white-, grey-, and black-box MIAs. Unlike conventional approaches that use a single query, MADreMIA utilizes chained generations - where each output informs the subsequent input - to amplify membership evidence. We demonstrate that training "re-members" exhibit significantly higher coherence and slower degradation during iterative regeneration than non-members generations. Our results across image, text, and audio modalities show that MADreMIA provides substantially richer signals for both membership and dataset inference across diverse model families, including IARs, diffusion, and large language models.