Differentially Private Fair Binary Classifications
Hrad Ghoukasian, Shahab Asoodeh
TL;DR
This work tackles binary classification under dual constraints of differential privacy and fairness, focusing on statistical parity as the fairness objective. It develops a two-stage framework: first training separate classifiers for each demographic group and then applying a post-processing step to enforce $Δ_{SP}(\hat{h})$ with minimal perturbation to the original predictions; it then extends this framework to provide $(ε',δ')$-DP guarantees using Laplace noise and DP-SGD-based base classifiers. Theoretical contributions include a lower bound on prediction changes under SP, a non-private algorithm (Algorithm 1) achieving perfect SP with optimal utility, and a private algorithm (Algorithm 2) delivering DP and SP guarantees along with average- and high-probability utility bounds. Empirically, the approach yields competitive accuracy with markedly improved fairness guarantees on the Adult and Credit Card datasets compared to DP-FERMI across multiple privacy accounting methods, demonstrating practical viability of DP+fairness in real-world tasks.
Abstract
In this work, we investigate binary classification under the constraints of both differential privacy and fairness. We first propose an algorithm based on the decoupling technique for learning a classifier with only fairness guarantee. This algorithm takes in classifiers trained on different demographic groups and generates a single classifier satisfying statistical parity. We then refine this algorithm to incorporate differential privacy. The performance of the final algorithm is rigorously examined in terms of privacy, fairness, and utility guarantees. Empirical evaluations conducted on the Adult and Credit Card datasets illustrate that our algorithm outperforms the state-of-the-art in terms of fairness guarantees, while maintaining the same level of privacy and utility.