Membership Inference Attacks on Discrete Diffusion Language Models

In three linesStudy of membership inference attacks (MIA) on masked diffusion language models (MDLM). Researchers extract 46-dimensional feature vectors from reconstruction loss at different masking ratios and train XGBoost and MLP classifiers. On MIMIR benchmark, XGBoost achieves AUC 0.878 (peak 0.930), outperforming SAMA baseline by 0.062 AUC. ELBO trajectory alone drives most of the signal.

Summary generated by Claude — human-verified