Auditing Reveals Fairness-Privacy Trade-offs at Subpopulation Level

This research provides the first in-depth investigation into the complex interplay between fairness-enhancing algorithms and privacy leakage in machine learning models. While previous work often focused on how privacy techniques affect fairness, this study reverses the perspective, analyzing the privacy implications of fairness interventions. A key contribution is the adaptation of the Likelihood Ratio Attack (LiRA) for subgroup auditing, which allows for the detection of privacy disparities that are obscured by aggregate-level evaluations. The findings reveal that the impact of fairness interventions on privacy risk is not uniform. Instead, it varies significantly based on factors such as the model's architecture, the size of specific subpopulations, and the chosen fairness mitigation strategy. The study also explores how Differential Privacy (DP) interacts with fairness methods, demonstrating that DP's benefits and costs are unevenly distributed across different subpopulations. This highlights the necessity of jointly evaluating fairness, privacy, and utility at a granular, subpopulation level, for which the authors introduce a unified empirical framework.