arXiv Computation and Language

Membership Inference in Fine-tuned Diffusion Language Models via Token-level Memorization Asymmetry

The paper introduces a new privacy vulnerability in diffusion language models (DLMs) called token‑level memorization asymmetry, derived from theoretical analysis of diffusion training dynamics. It proposes Q‑Skew, a quantile‑weighted skewness indicator, to perform membership inference on fine‑tuned DLMs, outperforming existing baselines across multiple datasets and models. Additionally, Q‑Skew can be used to extract personally identifiable information (PII), demonstrating a broader privacy attack surface.

arXiv AI
1d ago

Batch Normalization Amplifies Memorization and Privacy Risks

Batch Normalization (BN) is widely used to speed up and stabilize deep neural network training, yet its effect on privacy and memorization has been largely unexplored. This study shows that BN significantly increases the memorization of atypical or outlier samples, as evidenced by unintended memorization, per-sample influence, and heightened susceptibility to membership inference attacks across multiple datasets and architectures. A mechanistic analysis of the BN backward pass reveals that BN amplifies the per‑step margin growth of outlier samples during training, thereby intensifying their influence.

By Ngoc Phu Doan, Chongyan Gu, Ihsen Alouani
arXiv Machine Learning
Sep 11

Adaptive Diffusion Freezing: Privacy-preserving Diffusion Models Against Membership Inference Attacks

Adaptive Diffusion Freezing (ADF) is a new privacy‑preserving framework for diffusion models that protects against membership inference attacks (MIAs). It uses cross‑timestep adaptive freezing training, where a mask matrix controls which data subsets participate at each diffusion timestep, reducing over‑memorization and aligning model behavior for member and non‑member samples. A pretraining‑based risk‑aware freezing policy estimates MIA risk and suppresses high‑risk subset‑timestep pairs, achieving a superior privacy‑utility‑efficiency trade‑off across multiple datasets.

By Jialu Guo, Xiao Han, Junjie Wu
arXiv Machine Learning
Jun 2

Causal Evaluation of Membership Inference Attacks

arXiv:2602. 02819v4 Announce Type: replace Abstract: Membership Inference Attacks (MIAs) aim to distinguish training points (members) from unseen data (non-members), and are widely used to quantify memorization and assess privacy risks.

By Mathieu Even, Cl\'ement Berenfeld, Linus Bleistein, Tudor Cebere, Julie Josse, Aur\'elien Bellet
arXiv Machine Learning
Sep 11

Black-Box Membership Inference via Word-Level Probability Estimation

The paper introduces Word-level Probability MIA (WPMIA), a black-box membership inference attack that estimates word-level generation probabilities via Monte Carlo sampling and local kernel smoothing, then aggregates them into a sequence-level likelihood estimator. By conditioning on different prefixes, WPMIA amplifies distributional differences between member and non-member texts, outperforming existing black-box baselines on open-source LLMs and achieving an average TPR@5%FPR of 42.0 on proprietary models such as GPT‑5‑Chat, Gemini‑2.5‑Flash, and Claude‑4.5‑Haiku.

By Shengjie Niu, Yeheng Ge, Jian Huang
arXiv AI
Sep 2

The Privacy-Hallucination Tradeoff in Differentially Private Language Models

The paper investigates a privacy‑hallucination tradeoff in differentially private (DP) language models. It shows that DP pre‑training or fine‑tuning increases hallucinations compared to non‑DP models, especially as the privacy budget becomes stricter. The authors attribute this to DP mechanisms flattening output distributions, and demonstrate that controlling the frequency of facts in training data can mitigate hallucination risks.

By Krithika Ramesh, Krishna Pillutla, Danish Pruthi, Anjalie Field