LPF-IVN: A Lightweight Privacy Enhancing Scheme with Functional Mechanism of Intelligent Vehicle Networking

Abstract

Due to decentralization and effective prevention of privacy leakage, Differential Private Federated Learning(DP-FL) has emerged as an efficient technique in the Internet of Vehicles (IoV). However, the essence of key industrial is big data. When applying the DP-FL model to the IoV, these large-scale nonlightweight data such as Non-IID and high-dimensional will decrease the security and accuracy of the model. Therefore, for the security and accuracy of the IoV, we proposed a lightweight DP-FL framework called DPF-IVN, considering the impact of heterogeneous and privacy leak in the context of IoV. It adopts the idea of “lowering dimension first and then optimization” to process non-light quantified data in the IoV. Specifically, we novelly design a Federated Randomized Principal Component Analysis (FRPCA) algorithm, allowing users to map local data to low-dimensional data. Then, we propose the Functional Mechanism(FM) to disturb the gradient parameters to solve the problem of low training accuracy caused by gradient cutting. Besides, to reduce model differences, we used the Bregman dispersal as a regularized item update loss function to improve the accuracy of the model. Extensive experiments demonstrate the superior performance of DPF-IVN in the heterogeneous environment compared to other methods.