A new arXiv preprint introduces a membership inference method that estimates word-level probabilities to detect whether text appeared in a language model's training data. The approach targets black-box settings, where attackers lack direct access to model internals. It aims to improve privacy auditing of large language models.
A systematization of knowledge paper reviews 25 studies showing how explainable AI outputs can be turned against the models they describe. The authors group these attacks into model extraction, membership inference and model inversion, and note that explanations widen the confidentiality and privacy risks of deployed systems. The work calls for treating explanation interfaces as part of the attack surface.
A new paper introduces an auditing framework that runs membership inference attacks at the subgroup level against synthetic text produced under differential privacy. It explores whether formal worst-case privacy guarantees hold up in practice for smaller groups represented in the underlying data. The work offers data publishers a way to gauge real-world leakage before sharing synthetic text in place of sensitive datasets.