Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Efficient Federated Unlearning with Adaptive Differential Privacy Preservation

Yu Jiang, Xindi Tong, Ziyao Liu, Huanyi Ye, Chee Wei Tan, Kwok-Yan Lam

TL;DR

The proposed FedADP incorporates an adaptive differential privacy (DP) mechanism, carefully balancing privacy and unlearning performance through a novel budget allocation strategy tailored for FU, and employs a dual-layered selection process, focusing on global models with significant changes and client updates closely aligned with the global model, reducing storage and communication costs.

Abstract

Federated unlearning (FU) offers a promising solution to effectively address the need to erase the impact of specific clients' data on the global model in federated learning (FL), thereby granting individuals the ``Right to be Forgotten". The most straightforward approach to achieve unlearning is to train the model from scratch, excluding clients who request data removal, but it is resource-intensive. Current state-of-the-art FU methods extend traditional FL frameworks by leveraging stored historical updates, enabling more efficient unlearning than training from scratch. However, the use of stored updates introduces significant privacy risks. Adversaries with access to these updates can potentially reconstruct clients' local data, a well-known vulnerability in the privacy domain. While privacy-enhanced techniques exist, their applications to FU scenarios that balance unlearning efficiency with privacy protection remain underexplored. To address this gap, we propose FedADP, a method designed to achieve both efficiency and privacy preservation in FU. Our approach incorporates an adaptive differential privacy (DP) mechanism, carefully balancing privacy and unlearning performance through a novel budget allocation strategy tailored for FU. FedADP also employs a dual-layered selection process, focusing on global models with significant changes and client updates closely aligned with the global model, reducing storage and communication costs. Additionally, a novel calibration method is introduced to facilitate effective unlearning. Extensive experimental results demonstrate that FedADP effectively manages the trade-off between unlearning efficiency and privacy protection.

Efficient Federated Unlearning with Adaptive Differential Privacy Preservation

TL;DR

The proposed FedADP incorporates an adaptive differential privacy (DP) mechanism, carefully balancing privacy and unlearning performance through a novel budget allocation strategy tailored for FU, and employs a dual-layered selection process, focusing on global models with significant changes and client updates closely aligned with the global model, reducing storage and communication costs.

Abstract

Federated unlearning (FU) offers a promising solution to effectively address the need to erase the impact of specific clients' data on the global model in federated learning (FL), thereby granting individuals the ``Right to be Forgotten". The most straightforward approach to achieve unlearning is to train the model from scratch, excluding clients who request data removal, but it is resource-intensive. Current state-of-the-art FU methods extend traditional FL frameworks by leveraging stored historical updates, enabling more efficient unlearning than training from scratch. However, the use of stored updates introduces significant privacy risks. Adversaries with access to these updates can potentially reconstruct clients' local data, a well-known vulnerability in the privacy domain. While privacy-enhanced techniques exist, their applications to FU scenarios that balance unlearning efficiency with privacy protection remain underexplored. To address this gap, we propose FedADP, a method designed to achieve both efficiency and privacy preservation in FU. Our approach incorporates an adaptive differential privacy (DP) mechanism, carefully balancing privacy and unlearning performance through a novel budget allocation strategy tailored for FU. FedADP also employs a dual-layered selection process, focusing on global models with significant changes and client updates closely aligned with the global model, reducing storage and communication costs. Additionally, a novel calibration method is introduced to facilitate effective unlearning. Extensive experimental results demonstrate that FedADP effectively manages the trade-off between unlearning efficiency and privacy protection.

Paper Structure

This paper contains 29 sections, 1 theorem, 20 equations, 6 figures, 1 table, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Federated Unlearning
  4. Aggregation Methods
  5. Preliminaries
  6. Federated Learning
  7. Differential Privacy
  8. Threat Model
  9. Design of FedADP
  10. Overview
  11. Design of Adaptive DP and Dual-layered Selection in FL
  12. Adaptive Differential Privacy
  13. Dual-layered Selection
  14. Design of Calibration in FU
  15. Calibration of Historical Information
  16. ...and 14 more sections

Key Result

Theorem 1

To guarantee $(\epsilon,\delta)$-DP for all clients, the convergence upper bound after $T$ rounds is given by where $A= 1 - 2 \mu \eta + \mu \eta^2 L$.

Figures (6)

  • Figure 1: The workflow of FedADP from FL to FU. In FL, clients train local models by adaptively applying differential privacy while the server adopts model and update selection. In FU, the server calibrates historical information for unlearning.
  • Figure 2: Test accuracy of unlearning ratio of 5%, 10% and 25% on three different datasets of MNIST, CIFAR-10 and AGNews.
  • Figure 3: Test accuracy for different methods of Retrain, FedRecover, FedADP without DP and FedADP on dataset MNIST, CIFAR-10 and AGNews.
  • Figure 4: MISR and ASR for different methods of Retrain, FedRecover, FedADP without DP and FedADP on datasets MNIST, CIFAR-10 and AGNews.
  • Figure 5: Test accuracy under different secure aggregation methods of FedAvg, Trimmed Mean and Median for different methods of FedADP without DP and FedADP on three different datasets of MNIST, CIFAR-10 and AGNews.
  • ...and 1 more figures

Theorems & Definitions (3)

  • Definition 1: ($\epsilon$, $\delta$)-differential privacy dwork2006differential
  • Theorem 1
  • proof