UOIM: Online Incentive Design for Fair Hierarchical Federated Learning and Unlearning in Edge Computing

Abstract

Hierarchical Federated Learning (HFL) has emerged as a promising evolution of Federated Learning (FL) where a shared model is learned collaboratively via hierarchical aggregation, and it requires federated unlearning to protect users’ right to be forgotten in the training process. However, most existing designs neglect the unfair issue in federated unlearning, i.e., it is intuitively unfair for remaining users to spend additional costs to achieve the unlearning of target data of leaving users. This issue becomes even more serious in the dynamical HFL scenarios, where multi-level aggregation involves more intermediate users that may leave at any time, and the consecutive leaving of users will accumulate unfair costs for subsequent remaining users. Herein, we aim to bridge this gap via an unlearning-aware online incentive design. First, we design an online learning-based repeated auction to model the dynamics of the real-time participant selection process, where the multi-armed bandit machine is used to dynamically select participants in each round and assign the response reward. Moreover, we introduce a membership inference-based unlearning operation into the auction model for revoking target data of leaving users. Rigorous theoretical analysis and extensive experimental results show the effectiveness of the proposed mechanism and provide effective insights and strategies for the resource trading in the HFL system to complete the federated training.