A mutual information based federated learning framework for edge computing networks

Abstract

With the application of artificial intelligence in all field of life, people pay more attention to user privacy and data security. Under the condition of protecting user privacy, the federated learning model has become a popular research technology to solve the data islands problems. The edge computing network can be applied to smart city, Internet of vehicles and so on. Federated learning is a framework in which multiple hosts jointly learn a machine learning model. Each work device maintains the local model of its local training dataset, while the master device maintains the global model by aggregating the local models from the work devices. However, it cannot ensure that every local work device is an honest user because of a phenomenon that the hosts has been operated by attacker interferes in the process of local model training. In this paper, we assume that malicious nodes upload unreal learning parameters in the federated learning framework, which the global model will have high error rate. We propose a federated learning parameter aggregating algorithm based on mutual information. We introduced the relevance of model training learning rate to determine the consistency of the training direction of the local and central models at coarse granularity. We aggregated the parameters of the models at fine granularity based on the correlation of the gradients based on the mutual information. The mutual information method is used to calculate the similarity of the gradient trend between the local training model and overall model. We set the trust weight of each work device to reduce the negative impact of malicious nodes. The evaluation results show that the classification accuracy of the MIFL model is improved as compared with the average federated learning without malicious node. Especially, in the case of existing malicious nodes, the proposed algorithm can defend against malicious node attacks and sustain the robustness of Federated learning.