Rethinking Federated Unlearning via the Lens of Memorization

2026-05-23

AI Summary

This work addresses the limitations of existing federated unlearning methods, which often fail to completely erase target data and induce unfairness across clients due to neglecting overlapping information between data to be forgotten and retained data. From a memory-centric perspective, the authors propose removing only the unique memory attributable to the to-be-forgotten data while preserving shared knowledge supported by the remaining data. To this end, they introduce a grouped memory evaluation metric to distinguish between exclusive and shared memory components, and develop a federated memory pruning approach that integrates instance-level memory quantification, parameter pruning, and local model updating to precisely reset redundant memory parameters. Experiments demonstrate that the method effectively eliminates target memories while maintaining the utility of retained knowledge, achieving performance on par with full retraining baselines and significantly outperforming current federated unlearning algorithms.

Abstract

Federated learning (FL) increasingly needs machine unlearning to comply with privacy regulations. However, existing federated unlearning approaches may overlook the overlapping information between the unlearning and remaining data, leading to ineffective unlearning and unfairness between clients. In this work, we revisit federated unlearning through the lens of memorization. We argue that unlearning should mainly remove the unique memorized information attributable to the data to be forgotten, while preserving overlapping patterns that are also supported by the remaining data. Specifically, we propose Grouped Memorization Evaluation, an example-level metric that separates memorized knowledge from overlapping knowledge. Building on this metric, we introduce Federated Memorization Pruning (FedMemPrune), a pruning-based unlearning approach that resets redundant parameters responsible for memorization. Extensive experiments show that FedMemPrune closely matches retraining-based unlearning baselines while more effectively eliminating memorization than existing federated unlearning algorithms, yielding strong unlearning performance without sacrificing the utility of retained knowledge.

Problem

• Federated Unlearning
• Memorization
• Privacy Regulations
• Overlapping Information
• Fairness

Innovation

• federated unlearning
• memorization
• pruning
• overlapping knowledge
• privacy