JUMP Boosts Membership Inference on Diffusion Language Models.

Membership inference attacks (MIAs) are crucial for assessing data privacy in machine learning models, determining if specific data points were used in training. This research focuses on MIAs against fine-tuned discrete diffusion language models (dLLMs), which differ from autoregressive models by allowing attackers to mask arbitrary token sets and obtain parallel token distributions. Previous dLLM attacks, like SAMA, relied on multiple queries and random masking. The new attack, JUMP (Joint Uncertainty-Guided Mask Probing), leverages dLLMs' distinctive properties: any-order decodability to select low-reference-confidence positions, and parallel decodability to score these positions in a single joint masked query. JUMP computes a clipped target/reference reconstruction-gap statistic over these jointly masked positions. Evaluations on LLaDA-8B-Base across six domains show JUMP significantly outperforms SAMA, improving ROC-AUC from 0.82 to 0.90 and enhancing low-FPR detection. Crucially, JUMP achieves this with only one selector pass and one scoring pass per model, making it far more efficient.