Measuring Local and Shuffled Privacy of Gradient Randomized Response

Marin Matsumoto, Tsubasa Takahashi, Seng Pei Liew, Masato Oguchi

Keywords: Differential Privacy, Privacy Audit, Federated Learning

TL;DR: We introduce an empirical privacy test in FL clients by measuring empirical DP in local model and shuffle model.

Abstract:

Local differential privacy (LDP) provides a strong privacy guarantee in a distributed setting such as federated learning (FL).
Even if deployed LDP mechanisms honestly provide such privacy guarantees by randomizing gradients, how can we confirm and measure it?
To answer the above question, we introduce an empirical privacy test in FL clients by measuring the lower bounds of LDP.
The results of this measurement give the client the empirical $\epsilon$ and probability that the two gradients can be distinguished.
We then instantiate five adversaries in FL under LDP to measure empirical LDP at various attack surfaces, including a worst-case attack that reaches the theoretical upper bound of LDP.
The empirical privacy test with the adversary instantiations enables FL clients to understand LDP more intuitively and verify that mechanisms claiming $\epsilon$-LDP actually provide equivalent privacy protection.
We also demonstrate numerical observations of the measured privacy in these adversarial settings, and the randomization algorithm LDP-SGD is vulnerable to gradient manipulation and a well-pre-trained model.
We further discuss employing a shuffler to measure empirical privacy in a collaborative way and also measuring privacy of shuffled model.

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.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

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: 2925