Secure Aggregation With Logarithmic Overhead for Federated Learning in VANETs

Abstract

As a critical component in federated learning (FL), secure aggregation enables the server to learn the aggregated model without observing clients’ local training gradients. However, limited by computation and communication capabilities, existing aggregation schemes are not suitable to be directly employed in the Vehicular Ad Hoc Networks (VANETs) scenario. In this paper, we present a secure aggregation framework constructed with k-regular graph over VANETs scenario. We first optimize the secure aggregation scheme proposed by Bell et al. (CCS 2020). Specifically, using this new building block and an identity authentication mechanism in the vehicle-to-vehicle (V2V) communication mode, we design an optimized aggregation scheme that, when executed among n vehicles, can further reduce $2n$ communication times between vehicles and the central server while guaranteeing logarithmic overhead. Besides, by applying a zero-knowledge proof to the authentication process, our proposal supports vehicles anonymously constructing the k-regular graph and completing parameter computation process, which enhances privacy preservation in semi-honest settings. Under the experiment and security analysis, our proposal is demonstrated to be able to effectively achieve privacy preservation while achieving less computation and communication overheads compared to state-of-the-art aggregation schemes.