Statistical privacy audits are a standard way to check whether a differentially private mechanism lives up to its theoretical guarantees. Auditors typically run membership inference attacks (MIAs) to estimate a lower bound on the true privacy loss. A new arXiv preprint (2610.27021v1) draws attention to a complication in this process: selection bias.
The paper's title and abstract indicate that biased selection of data or audit conditions can undermine the reliability of these estimates. If the audit setup is not representative, the estimated lower bound may not reflect the mechanism's actual privacy behavior. The work therefore targets reliable estimation of differential privacy under selection bias.
Because the abstract is brief, the specific method is not described in detail here. The main contribution, as stated, is addressing a gap in how privacy audits are conducted and interpreted when selection bias is present.