Communication-Efficient Federated Learning Based on Chained Secure Multiparty Computing and Compressed Sensing
Di Xiao, Pengcen Jiang, Min Li · 2022
Federal learning (FL), as a new technology with privacy protection capabilities, has been widely used. However, the protection capabilities of FL are limited and a lot of communication overhead is required, which is often limited by communication environments. In order to improve the privacy protection and reduce the communication burden, we combine secure multi-party computing(SMC) and compressed sensing(CS) to propose a new chained federated learning scheme called Chain-FL via CS (Chain-FL-CS). We first design a heuristic compression transmission strategy for the model based on CS so that the model can be compressed with a lower loss. Then we use a chained-SMC to perform lossless and secure aggregation of the calculation results. This technology mainly uses masking mechanism for communication. Compared with other schemes, ours has lower complexity. We have conducted experiments using a Deep Neural Network Model (DNN). Experimental results show that our scheme can guarantee the convergence of model under the premise of high transmission efficiency, and it has certain privacy protection capability.