FedMMD: A Federated Weighting Algorithm Considering Non-IID and Local Model Deviation

Federated learning (FL) is a distributed machine learning method to protect users' privacy and security. It currently faces the following two problems (McMahan et al., 2017): (1) The performance of the global model degrades when dealing with Non-Independent Identically Distributed (Non-IID) data. Other existing classical methods do not have a rigorous theory based on error processing ; (2) In the process of global model aggregation, the classical FL algorithm either directly averages local models or solely considers the proportion of local data to assign weights to the models, without accounting for the discrepancies between the local models. In this paper, a federated aggregation algorithm Federated Maximum Mean Discrepancy (FedMMD) is proposed to address the deviation between the local models and Non-IID. First of all, this paper utilizes the Dilated Convolution Meet Transformer (DCMT) model for local model feature extraction. This approach aims to capture more feature information and minimize the impact of Non-IID scenarios. Secondly, the learning stability of the local model data participating in the aggregation is compared with Maximum Mean Discrepancy (MMD). The weights of the local models involved in the aggregation are determined using the SKNQ (Student-Keuls-Newman-Q) method and the entropy weight method. The SKNQ method calculates and compares the MMD across multiple clients, while the entropy weight method assigns different weights to clients based on their deviations. On the standard dataset, the FedAvg algorithm (McMahan et al., 2017), the FedProx algorithm (Li et al., 2020) and the FedMMD algorithm are compared. The experimental results demonstrate that the FedMMD algorithm, used in training the global model, not only enhances the accuracy of learning IID and Non-IID data but also improves the generalization capability of the global model.