A new arXiv paper examines how to audit user-level privacy guarantees in Private Evolution, a method for generating synthetic data in federated settings. The approach collects clipped user votes over a shared candidate bank and turns them into a differentially private histogram with calibrated noise. The work focuses on verifying that individual users' raw data remains protected under this mechanism.
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 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.