Differentially-private and plausible counterfactuals

Abstract: Counterfactual explanations are particularly appealing in high-stakes domains such as finance and hiring, as they provide affected users with suggestions on how to alter their profiles to receive a favorable outcome. However, existing methods are characterized by a privacy-quality trade-off. More precisely, as highlighted in recent works, instance-based approaches generate plausible counterfactuals but are vulnerable to privacy attacks, while perturbation-based methods offer better privacy at the cost of lower explanation quality. In this paper, we propose to solve this dilemma by introducing a diverse set of differentially-private mechanisms for generating counterfactuals, providing a high resistance against privacy attacks while maintaining high utility. These mechanisms can be integrated at different stages of the counterfactual generation pipeline i.e, pre-processing, in-processing or post-processing), thereby offering maximal flexibility during the design for the model provider. We have performed an empirical evaluation of the proposed approaches on a wide range of datasets and models to evaluate their effect on the privacy and utility of the generated counterfactuals. Overall, the results obtained demonstrate that in-processing methods significantly reduce the success rate of privacy attacks while moderately impacting the quality of counterfactuals generated. In contrast, pre-processing and post-processing mechanisms achieve a higher level of privacy but at a greater cost in terms of utility, thus being more suitable for scenarios in which privacy is paramount.

Submission Type: Long submission (more than 12 pages of main content)

Changes Since Last Submission: We addressed the reviewers' comments in the new version. Supplementary materials include updated complementary results in accordance with the new revision, in addition to the results for one more dataset.

Assigned Action Editor: ~Sanghamitra_Dutta2

Submission Number: 6484