Lower Bounds on Post-Audit Manipulation in Fairness Auditing

A recent study published on arXiv (2608.00568) establishes theoretical minimum limits on the potential manipulation by companies following a fairness audit, even when the auditor's resources are limited. This research conceptualizes fairness auditing as a min-max optimization problem involving an unlimited company and a resource-restricted auditor. Two types of auditing frameworks are examined: one in which the auditor uses a predetermined audit set to certify fairness, and another that necessitates the audit set to achieve fairness estimation within an alpha approximation. For both scenarios, the study provides explicit lower bounds regarding the maximum demographic parity deviation post-audit, based on the audit budget. These results enhance previous findings on the limitations of black-box fairness auditing, indicating that sufficiently complex models can bypass any auditing methods. The implications of this research are significant for critical areas such as hiring, lending, and automated decision-making, where fairness audits are becoming increasingly necessary. The paper is authored by a team of researchers and was noted as a cross-type submission on arXiv.