Federated Unlearning in the Wild: Rethinking Fairness and Data Discrepancy

TL;DR

This work investigates federated unlearning under realistic, cross-domain data distributions and fairness constraints, arguing that existing exact retraining and approximate unlearning methods overestimate effectiveness due to unfair cross-client costs and synthetic benchmarks. It introduces Cross-Domain FU Benchmark to evaluate methods under Real-Noniid conditions, revealing limitations of current approaches. A novel method, Federated Cross-Client-Constrains Unlearning (FedCCCU), identifies key neurons via local attribution and applies Rank-Based Selection to suppress forgetting-relevant components while preserving non-forgetting clients, achieving better forgetting with fewer collateral effects. The study emphasizes the need for fairness-aware, real-world FU systems and provides a principled framework and empirical evidence to guide future research toward deployable, responsible federated unlearning.

Abstract

Machine unlearning is critical for enforcing data deletion rights like the "right to be forgotten." As a decentralized paradigm, Federated Learning (FL) also requires unlearning, but realistic implementations face two major challenges. First, fairness in Federated Unlearning (FU) is often overlooked. Exact unlearning methods typically force all clients into costly retraining, even those uninvolved. Approximate approaches, using gradient ascent or distillation, make coarse interventions that can unfairly degrade performance for clients with only retained data. Second, most FU evaluations rely on synthetic data assumptions (IID/non-IID) that ignore real-world heterogeneity. These unrealistic benchmarks obscure the true impact of unlearning and limit the applicability of current methods. We first conduct a comprehensive benchmark of existing FU methods under realistic data heterogeneity and fairness conditions. We then propose a novel, fairness-aware FU approach, Federated Cross-Client-Constrains Unlearning (FedCCCU), to explicitly address both challenges. FedCCCU offers a practical and scalable solution for real-world FU. Experimental results show that existing methods perform poorly in realistic settings, while our approach consistently outperforms them.

Figures (4)

• Figure 1: Federated unlearning in single-domain (top) and cross-domain (bottom) settings.
• Figure 2: The distribution of data feature characteristics of the client in different allocation scenarios
• Figure 3: FedCCCU: Federated Cross-Client-Constrains Unlearning
• Figure 4: Performance Comparison of Different Unlearning Strategies on an Image Recognition Task