Federated Learning with Model Pruning in Resources-Constrained Mobile Edge Networks

To minimize Federated Learning (FL) training overhead in mobile edge networks, this paper proposes a FL scheme with Online joint model Pruning and resource Allocation (FLOPA). First, we formulate the optimization problem with the goal of minimizing long-term training overhead. Second, considering the energy constraints of clients, we transform the problem of minimizing training overhead into an online per-round model pruning and resource allocation problem based on Lyapunov theory. Finally, we decompose the problem into three subproblems and propose an iterative algorithm that combines convex optimization and block coordinate descent. Experimental results show that FLOPA outperforms existing benchmark schemes in various scenarios.