Federated Learning with non‐IID data in Mobile Edge Computing Systems
Chenyuan Feng, Daquan Feng, Zhongyuan Zhao, Geyong Min, Hancong Duan · 2023
Most of the existing works with respect to federated learning (FL) are mainly based on independent and identically distributed (IID) data; however, it is still unknown how to improve the performance of FL with non-IID data in the MEC systems. Compared with centralized learning with IID data, we analyze the weight difference of models trained by it and the FL with non-IID data. We also derive a theoretical upper bound and redesign the federated averaging scheme to reduce the weight difference. To further mitigate the impact of non-IID data, we design a data sharing scheme based on our formulated optimization problem which jointly minimize the accuracy loss and the energy consumption and latency with constrained resource of MEC systems. Then we propose a computation-efficient algorithm to approach the optimal solution and provide the experiment results to evaluate our proposed schemes. Our theoretical upper bound is verified as tight. Moreover, our proposed federated averaging and data sharing scheme can significantly enhance the test accuracy with non-IID dataset especially for small user cases, and our optimization resource allocation and user selection scheme help greatly reduce energy consumption and latency. The experimental results show that by using our proposed algorithm, the performance of FL with non-IID data cases can approach the centralized leaning with IID data case in a cost-efficient way.