Continual Learning with Differential Privacy

TL;DR

This paper introduces a new method for continual learning that preserves differential privacy, ensuring data privacy across tasks while maintaining model performance, through a novel algorithm and theoretical analysis.

Contribution

We propose a differential privacy-preserving algorithm for continual learning that bounds privacy loss using a new notion of continual adjacent databases and a moments accountant.

Findings

The algorithm provides formal privacy guarantees across tasks.

It tightens privacy loss compared to existing methods.

Preliminary results show maintained model utility.

Abstract

In this paper, we focus on preserving differential privacy (DP) in continual learning (CL), in which we train ML models to learn a sequence of new tasks while memorizing previous tasks. We first introduce a notion of continual adjacent databases to bound the sensitivity of any data record participating in the training process of CL. Based upon that, we develop a new DP-preserving algorithm for CL with a data sampling strategy to quantify the privacy risk of training data in the well-known Averaged Gradient Episodic Memory (A-GEM) approach by applying a moments accountant. Our algorithm provides formal guarantees of privacy for data records across tasks in CL. Preliminary theoretical analysis and evaluations show that our mechanism tightens the privacy loss while maintaining a promising model utility.