Linus Bleistein

Tudor Cebere, Mathieu Even, Linus Bleistein et al.

Privacy auditing provides empirical lower bounds on the differential privacy parameters of learning algorithms. Existing methods, however, require interventional access to the training pipeline, either to retrain multiple times or to randomize data inclusion. This is often infeasible for large deployed systems such as foundation models. We introduce Zero-Run privacy auditing, a post-hoc framework for auditing models using two fixed datasets: examples known to be training-set members and examples known to be non-members. In this observational regime, membership is no longer randomized; instead, member and non-member data often differ in distribution, so membership inference scores may reflect a distribution shift rather than algorithmic leakage. Drawing on ideas from causal inference, we formalize this confounding effect and propose two complementary corrections that yield valid privacy audits. Our first approach models the combined effect of distribution shift and algorithmic leakage as an adaptive composition, producing conservative global corrections. Our second approach conditions on observed data and adjusts pointwise membership guesses, yielding sharper instance-dependent bounds. Experiments on synthetic data and large-scale models show that Zero-Run auditing enables practical privacy evaluation when retraining or controlled data insertion is infeasible.

Mathieu Even, Clément Berenfeld, Linus Bleistein et al.

Membership Inference Attacks (MIAs) are widely used to quantify training data memorization and assess privacy risks. Standard evaluation requires repeated retraining, which is computationally costly for large models. One-run methods (single training with randomized data inclusion) and zero-run methods (post hoc evaluation) are often used instead, though their statistical validity remains unclear. To address this gap, we frame MIA evaluation as a causal inference problem, defining memorization as the causal effect of including a data point in the training set. This novel formulation reveals and formalizes key sources of bias in existing protocols: one-run methods suffer from interference between jointly included points, while zero-run evaluations popular for LLMs are confounded by non-random membership assignment. We derive causal analogues of standard MIA metrics and propose practical estimators for multi-run, one-run, and zero-run regimes with non-asymptotic consistency guarantees. Experiments on real-world data show that our approach enables reliable memorization measurement even when retraining is impractical and under distribution shift, providing a principled foundation for privacy evaluation in modern AI systems.