Auditing Fairness-Privacy Trade-offs: Subpopulation-Level Effects of Fairness-Enhancing Algorithms

This paper investigates the intersection of fairness and privacy in machine learning, specifically asking how fairness-enhancing algorithms influence membership inference privacy risks at the subpopulation level. While prior research has extensively studied how privacy-preserving techniques (like Differential Privacy) affect fairness, the inverse relationship—how fairness interventions affect privacy leakage—remains largely unexplored. The authors aim to determine if fairness interventions inadvertently increase privacy risks for specific demographic groups.

The researchers introduce a unified empirical framework to audit fairness-privacy trade-offs. They adapt the Likelihood Ratio Attack (LiRA) and other membership inference attacks (OQTA, OTA) to perform subgroup-specific auditing. By filtering datasets and attack distributions to specific subpopulations, they can measure privacy leakage with high granularity. The study evaluates five common fairness mitigators—including pre-processing (DIR, REW, SYN), in-processing (EGR), and post-processing (CPP) methods—across multiple model families and datasets. They also analyze how these interventions interact with Differential Privacy (DP) to understand the resulting trade-offs in utility and privacy.

The study demonstrates that fairness interventions do not have a uniform effect on privacy. Instead, the impact is highly contingent on the interaction between the model architecture, the representation of the subgroup in the training data, and the specific mitigation strategy employed. A key finding is that while Differential Privacy reliably reduces overall privacy risk, its utility costs are distributed unevenly across subpopulations, often causing performance to collapse for underrepresented groups. The authors conclude that fairness and privacy are not inherently in conflict, but they must be evaluated together at the subpopulation level to avoid masking significant disparities that aggregate metrics fail to capture.