Joint Optimizing Federated Learning and Unlearning over Wireless Networks

Abstract

In this paper, a communication and energy-efficient federated learning (FL) and unlearning (FU) co-designed framework is proposed. In the considered model, multiple edge devices collaboratively train a model. Users are allowed to quit the FL training and remove their data impacts on the trained model. Hence, an FU training is performed after the FL training. To reduce the energy and time consumption of FU, the users are grouped into multiple clusters such that the users in a group will retrain a deep learning model when the users in the group request to quit the FL training. The base station (BS) needs to determine the number of clusters, the users in each cluster, and the resource block (RB) allocated to each user, according to the probability that each user requests to quit the FL, data distribution, wireless channel conditions, and time and energy constraints. This problem is formulated as an optimization problem whose goal is to minimize the expected loss by determining the number of clusters, user cluster assignment, and RB allocations. This problem is challenging to solve since 1) the BS cannot predict which users will request data deletion in advance and 2) no explicit equation exists to capture the relationship between the number of clusters, user cluster assignment, and the expected FU loss. To solve the formulated problem, we first analyze the convergence rate of FU to build an explicit relationship among the FU training loss, the number of clusters, and the users in each cluster. Based on this relationship, the original problem is simplified to a mixed-integer optimization problem which can be divided into the user cluster assignment and the RB allocation optimization subproblems. These two problems are alternately solved by the designed Dinkelbach's method and the fast iterative shrinkage-thresholding algorithm (Dinkelbach-FISTA) and the standard Hungarian algorithm. Simulation results show that our method improves classification accuracy by up to 11.1% and achieves up to 3× faster convergence compared to gradient-based and ISTA-based methods.