Privacy leaks in natural language are rarely a single obvious identifier. The new arXiv paper, QuasiLeak-Bench, focuses on quasi-identifiers: pieces of information that seem harmless in isolation but, when grouped, can jointly expose an individual's private attributes. The authors call such a set of cues a quasi-evidence group.
The paper argues that existing privacy detectors may not be adequately tested against this kind of threat. To address that gap, the authors introduce QuasiLeak-Bench, a benchmark for evaluating LLM-based privacy detectors on what they term policy-conditioned quasi-evidence group recovery. The task asks whether a detector can reconstruct the full group of cues that, under a given policy, would constitute a privacy violation.
Because the abstract is truncated, details on the benchmark's construction and results are not yet available from the source. The key contribution, as stated, is framing privacy detection as a group-recovery problem rather than a simple classification of individual identifiers.