FedCSA: Boosting the Convergence Speed of Federated Unlearning under Data Heterogeneity

The exponential growth of the Internet and the widespread availability of personal data have raised significant concerns regarding personal privacy and societal security. In response, the “right to be forgotten” has emerged as a fundamental principle, empowering individuals to request the removal of their personal data. Though relieving the privacy concerns at a certain degree by doing so, Artificial Intelligence (AI) models, trained by personal data, still suffers from privacy leakage risks. Therefore, machine unlearning attracts great attention from both industry and academia, aiming at erasing the risks of privacy leakage of AI models. However, existing machine unlearning methods primarily operate within centralized frame-works, limiting their applicability to decentralized Federated Learning (FL) settings where data is decentralized and not directly accessible. Additionally, current federated unlearning methods only approximate data deletion, leaving a potential risk of data leakage. To address these issues, this paper presents FedCSA, The FedCSA approach adopts a divide-and-conquer strategy by clustering and slicing the users’ data. Subsequently, submodels are trained on each individual slice, and their aggregation yields the global model. During the unlearning task, only the corresponding sub-model undergoes retraining, significantly reducing training time while ensuring data privacy and achieving precise data deletion. Extensive experiments are conducted on real-world datasets, including MNIST, Fashion-MNIST, and CIFAR-10, to evaluate the performance of FedCSA. The results demonstrate that FedCSA outperforms state-of-the-art machine learning methods, such as retrain and SISA, in terms of efficiency and global model performance. Our code is publicly available at https://github.com/ZhenWang9/FedCSA.