Federated Unlearning via Memorization Pruning
A new research paper proposes FedMemPrune, a pruning-based federated unlearning method that removes memorized information unique to forgotten data while preserving overlapping patterns. The approach uses Grouped Memorization Evaluation to separate memorized from overlapping knowledge. Experiments demonstrate effectiveness in complying with privacy regulations.
Key facts
- Federated learning needs machine unlearning for privacy regulations.
- Existing approaches overlook overlapping information between unlearning and remaining data.
- FedMemPrune resets redundant parameters responsible for memorization.
- Grouped Memorization Evaluation is an example-level metric.
- The paper is on arXiv with ID 2605.24545.
- The method preserves overlapping patterns supported by remaining data.
- Extensive experiments show FedMemPrune's effectiveness.
- The work revisits federated unlearning through memorization.
Entities
Institutions
- arXiv