FedPCGA: A Federated Unlearning Method Based on Projected Conflict Gradient Ascent

As data becomes increasingly valuable, the importance of data privacy also grows. Governments and legislators worldwide have established regulations to safeguard users’ “Right to Be Forgotten”, which allows users to request the deletion of their data from training datasets. However, merely deleting the data is insufficient, as models may retain implicit information or knowledge about the users’ data acquired during training. Therefore, it is crucial to enable models to specifically forget targeted data information. In Federated Learning, where multiple participants and incremental training are involved, existing centralized machine unlearning methods are not directly applicable. In this paper, we propose a federated unlearning method based on projected conflict gradient ascent that removes the impact of class data from the trained global model. Since the unlearning process conflicts with the gradient of the original training task, we project the conflicting unlearning gradient onto the gradient of the final training round in the target client model, thereby avoiding interference with model performance. Additionally, post-processing approach are proposed to mitigate catastrophic forgetting while maintaining overall model performance. The experimental results demonstrate that our method obtains a performance comparable to federated retraining from scratch in accomplishing the unlearning task, with significantly reduced execution time.