NDSS 2025 – Diffence: Fencing Membership Privacy With Diffusion Models

This NDSS session paper introduces Diffence, a generative-model-based pre-inference defense that re-generates inputs to mitigate membership inference attacks without modifying the target model or degrading accuracy. Experiments across datasets show meaningful reductions in attack accuracy (~15.8%) and AUC (~14.0%) on undefended models, with additional gains when combined with defenses like SELENA, all while preserving prediction labels and confidence vector utility and adding only ~57 ms per sample.