Antipodes of Label Differential Privacy: PATE and ALIBI

Mani Malek Esmaeili, Ilya Mironov, Karthik Prasad, Igor Shilov, Florian Tramer

Keywords: differential privacy, label differential privacy, PATE, ALIBI, memorization attacks

TL;DR: We propose two very distinct approaches for training models with label-only differential privacy (Label DP) and evaluate them with memorization attacks.

Abstract:

We consider the privacy-preserving machine learning (ML) setting where the trained model must satisfy differential privacy (DP) with respect to the labels of the training examples. We propose two novel approaches based on, respectively, the Laplace mechanism and the PATE framework, and demonstrate their effectiveness on standard benchmarks.

While recent work by Ghazi et al. proposed Label DP schemes based on a randomized response mechanism, we argue that additive Laplace noise coupled with Bayesian inference (ALIBI) is a better fit for typical ML tasks. Moreover, we show how to achieve very strong privacy levels in some regimes, with our adaptation of the PATE framework that builds on recent advances in semi-supervised learning.

We complement theoretical analysis of our algorithms' privacy guarantees with empirical evaluation of their memorization properties. Our evaluation suggests that comparing different algorithms according to their provable DP guarantees can be misleading and favor a less private algorithm with a tighter analysis.

Code for implementation of algorithms and memorization attacks is available from https://github.com/facebookresearch/label_dp_antipodes.

Code Of Conduct: I certify that all co-authors of this work have read and commit to adhering to the NeurIPS Statement on Ethics, Fairness, Inclusivity, and Code of Conduct.

Code: https://github.com/facebookresearch/label_dp_antipodes/

Community Implementations: [ 1 code implementation](https://www.catalyzex.com/paper/arxiv:2106.03408/code)