Privacy-Preserving Algorithmic Recourse

Privacy risks exist in recourse explanations: The paper demonstrates that counterfactual explanations—which tell individuals how to change their circumstances to get better outcomes from AI systems—can expose sensitive personal information, and this risk increases when providing multi-step recourse paths rather than single recommendations.

Simple noise addition isn't enough: Adding random noise to recourse paths, a common privacy approach, makes the suggested changes less realistic and actionable for people trying to follow them. The paper shows that more sophisticated privacy methods are needed to maintain both privacy and usefulness.

PrivRecourse offers a practical solution: The proposed method uses differentially private clustering to create a privacy-safe map of data patterns, then generates recourse paths by connecting these protected data clusters. This approach balances two competing goals: keeping personal data private while giving people realistic, step-by-step guidance on how to improve their outcomes.

Tested on real financial decisions: The research was evaluated using finance datasets, suggesting the approach is relevant for high-stakes domains like loan approvals and credit decisions where both privacy protection and actionable recourse are important to individuals facing adverse outcomes.

Privacy-preserving recourse is achievable: The empirical results indicate that PrivRecourse can generate recourse paths that are simultaneously private (meeting differential privacy standards) and realistic enough for people to actually follow, addressing a key tension in responsible AI systems.