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.
Introduction: The Challenge of Forgetting in Federated Learning
Machine unlearning has emerged as a critical technical requirement to satisfy legal mandates like the General Data Protection Regulation (GDPR), which grants individuals the "right to be forgotten." In the context of Federated Learning (FL), where a global model is collaboratively trained across decentralized clients, unlearning becomes particularly complex. The standard "gold standard" for unlearning is to retrain the entire global model from scratch, excluding the data to be forgotten. However, this is computationally prohibitive and often unnecessary for clients whose data remains unchanged.
The authors identify two primary shortcomings in current Federated Unlearning (FU) research. First is the Fairness Dilemma: exact unlearning via retraining imposes a heavy burden on uninvolved clients, while approximate methods often lead to "collateral damage," where the model's performance on retained data degrades significantly. Second is the Pseudo-Noniid Fallacy: most existing benchmarks rely on synthetic data distributions (e.g., Dirichlet sampling) that do not reflect the true heterogeneity of real-world data, where different clients may operate in entirely different domains.
Rethinking Data Heterogeneity: The Real-Noniid Benchmark
Current FU methods are typically evaluated under IID (Independent and Identically Distributed) or "Pseudo-Noniid" settings. In these scenarios, clients share the same underlying dataset but receive different slices or class distributions. The authors argue that this fails to capture the "wild" reality of FL, where clients might have data from completely different sources—for instance, one client using MNIST (clean digits) while another uses SVHN (cluttered street-view numbers).
To address this, the paper introduces a "Real-Noniid" benchmark. This benchmark utilizes diverse datasets for different clients (e.g., MNIST, SVHN, and USPS for digit recognition) while maintaining overlapping label spaces to ensure task alignment. This setup reveals that many methods which appear effective on synthetic data fail when faced with significant feature shifts like changes in resolution or lighting.
Proposed Method: Federated Cross-Client-Constrains Unlearning (FedCCCU)
To overcome the limitations of retraining and blunt approximate methods, the authors propose FedCCCU. This method shifts the focus from global retraining to precise "model editing" by identifying and removing specific knowledge at the neuron level.
The core of FedCCCU is the identification of "dominant neurons." The process involves:
- Neuron Importance Calculation: Each client uses attribution-based scores to determine which neurons are most critical for the class they wish to forget.
- Dominant Score Computing: The system calculates a "Dominant Score" ($R$) for each neuron by comparing its importance for the "forgetting" client against its importance for all "retaining" clients. A low $R$ score indicates a neuron that is vital for the data to be deleted but largely irrelevant to other clients.
- Rank-Based Zeroing: The server identifies the top $n$ dominant neurons and sets their weights to zero. This surgically erases the target memory while preserving the shared knowledge essential for other clients.
Key Results and Insights
Experimental results demonstrate that FedCCCU effectively balances unlearning precision with cross-client fairness. In image recognition tasks, the method successfully reduced target class accuracy on the forgetting client (from ~94% to ~16-20%) while maintaining high performance for others. Crucially, FedCCCU mitigated the side effects common in other "zeroing" methods, which often cause catastrophic accuracy drops across the entire network.
An interesting insight from the study is that unlearning can sometimes act as a form of regularization. In highly heterogeneous settings, removing "conflicting" gradients from a specific client's data actually improved performance for some non-requesting clients. Furthermore, the efficiency of FedCCCU is a major advantage; by using weight pruning instead of retraining, it avoids the massive communication and computation overhead associated with traditional "Delete-Retrain" strategies.
Conclusion
FedCCCU represents a significant step toward practical and fair Federated Unlearning. By moving beyond synthetic benchmarks and addressing the "collateral damage" of approximate methods, it offers a scalable solution for real-world applications. Future work will look into extending this "dominant neuron" approach from class-level to sample-level unlearning and testing its robustness across non-visual data domains.
Summary updated