Practical Federated Unlearning: A Target Client-Driven Approach to Model Forgetting

To remove the contribution of specific data from the global model in federated learning, federated unlearning has recently emerged. Existing approaches face challenges such as reliance on full client participation, the need to store historical model updates, and high communication costs. To overcome these limitations, we propose Practical Federated Unlearning (PFU), a target client-driven approach to model forgetting with only a single round of interaction between the target client and the aggregation server during the unlearning stage. To fulfill PFU, the aggregation server utilizes the Fisher Information Matrix to identify the most sensitive parameters with respect to the target client’s data and sends their indices to the target client. Then, the target client estimates its contribution to the sensitive parameters using a Bayesian inference-based method, thus avoiding the requirement for data or historical model updates from all other clients. Finally, it prunes the sensitive parameters with a tailored mechanism, which achieves effective unlearning of the data knowledge while preventing catastrophic forgetting. Theoretical analysis under explicit regularity assumptions and experimental results confirm the effectiveness and practical advantages of PFU.