Differentially Pivate Per-Instance Additive Noise Mechanism: A Game Theoretic Approach
Primary Area: societal considerations including fairness, safety, privacy
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: differential privacy, safety, privacy
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: We introduce a per-instance noise variance optimization (NVO) game, demonstrating that its Nash equilibrium points provide guarantees of differential privacy while preserving data statistics more effectively than the conventional Laplace mechanism.
Abstract: Recently, the concept of per-instance differential privacy (pDP) has gained significant attention by virtue of its capability to assess the differential privacy (DP) of individual data instances within a dataset.
Traditional additive mechanisms in the DP domain, which add identical noises to all data instances, often compromise the dataset's statistical utility to guarantee DP.
A main obstacle in devising a per-instance additive noise mechanism stems from the interdependency of the additive noises: altering one data instance inadvertently affects the pDP of others.
This intricate interdependency complicates the problem, making it resistant to straightforward solutions.
To address this challenge, we propose a per-instance noise variance optimization (NVO) game, framed as a common interest sequential game.
We show that the Nash equilibrium (NE) points of this game inherently guarantee DP.
We leverage two algorithms to derive strategies for achieving the NE: 1) an approximate enumeration (AE) using a genetic algorithm, and 2) best response dynamics (BRD).
To validate the efficacy of our approach, we evaluate the NVO game on various statistical metrics including regression experimental results.
The source code to reproduce the results will be available soon.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 2502
Loading