Federated learning implementation based on DDS data distribution service

Abstract

Federated learning allows multiple local private clients to collaborate on training the same model without sharing each client's private data. It can achieve collaborative training between users while protecting data privacy and security. Therefore, it is advocated for use in the fields of Internet of Things, Internet of Vehicles, etc. There are security issues in the transmission process of federated learning models. How to ensure data privacy during local data collection, local model parameter transmission, model aggregation and performance testing while maximizing the use of private data is a question that people have been exploring. This paper proposes a DDS-based FL framework (FL-DDS) to achieve secure transmission of model parameters. The client and server are used as nodes in the DDS security domain, and the DDS data distribution mechanism is used to transmit model parameters. Through the DDS topic-based publish-subscribe mechanism, the global model and local model are transmitted through the DDS topic. At the same time, the DDS authentication component, access control component and encryption component are used to achieve domain-level security and intra-domain security of model transmission. Experiments show that (FL-DDS) can protect the privacy of model parameters without affecting communication performance.