Federated Unlearning: a Perspective of Stability and Fairness
Jiaqi Shao, Tao Lin, Xuanyu Cao, Bing Luo
TL;DR
This work analyzes federated unlearning (FU) under data heterogeneity, introducing three metrics—FU verification $V$, global stability $S$, and local fairness $Q$—to quantify unlearning side effects. It provides a theoretical framework showing fundamental trade-offs between verification, stability, and fairness, with bounds that depend on heterogeneity, gradient corrections, and divergence between groups. To balance these trade-offs, the authors design optimization-based FU mechanisms that incorporate penalty terms and gradient correction, plus fairness-constrained saddle-point formulations, accompanied by convergence guarantees. The framework includes practical FU algorithms operating in rounds with local updates and global gradient corrections, and is empirically validated on non-IID MNIST/CIFAR-10 to demonstrate improved stability and fairness relative to retraining, while preserving verification accuracy. Overall, the paper offers a principled, heterogeneity-aware approach to FU with theoretical guarantees and practical guidance for deploying verifiable, fair, and stable FU in federated systems.
Abstract
This paper explores the multifaceted consequences of federated unlearning (FU) with data heterogeneity. We introduce key metrics for FU assessment, concentrating on verification, global stability, and local fairness, and investigate the inherent trade-offs. Furthermore, we formulate the unlearning process with data heterogeneity through an optimization framework. Our key contribution lies in a comprehensive theoretical analysis of the trade-offs in FU and provides insights into data heterogeneity's impacts on FU. Leveraging these insights, we propose FU mechanisms to manage the trade-offs, guiding further development for FU mechanisms. We empirically validate that our FU mechanisms effectively balance trade-offs, confirming insights derived from our theoretical analysis.