Privacy Protection Technology of Maritime Multi-agent Communication Based on Part-Federated Learning

Abstract

Federated learning is a machine learning model based on distributed data sets, which builds a global model under the premise of ensuring the privacy and data security of participants. Because of this characteristic, federated learning is very suitable for maritime communication systems with large amounts of distributed data. However, the data set of maritime multi-agent communication system is different from the general data set, and the data distribution is not uniform, which increases the deviation of the model. In this paper, we propose a Part-Federated Learning (PFL) method which combines the advantages of split learning to improve the classical federated learning. This method, only uploading some parameters in the local model to the cloud server as shared parameters, reduces the communication cost of distributed learning, improves the privacy of the algorithm to the data, and has better performance in processing non-IDD distributed data. We optimize the proportion of shared parameters of PFL by considering the convergence of the algorithm and the communication cost. Finally, we verify the advantages of the algorithm in processing non-IID data through experiments, simulate the process of parameter optimization, and prove the feasibility of the algorithm.