BV-ICVs: A privacy-preserving and verifiable federated learning framework for V2X environments using blockchain and zkSNARKs

Abstract

As part of vehicle to everything (V2X) environments, intelligent connected vehicles (ICVs) generate a large amount of data, which can be exploited securely and effectively through decentralized techniques such as federated learning (FL). Existing FL systems, however, are vulnerable to attacks and barely meet the security requirements for real-world applications. If malicious or compromised ICVs upload inaccurate or low-quality local model updates to the central aggregator, they may reduce the accuracy of the global model, thereby reducing drivers safety and efficiency. This paper aims to alleviate these concerns by presenting BV-ICVs, a blockchain-enabled and privacy-preserving FL framework for ICVs in an edge-envisioned V2X environment. This system uses Zero-Knowledge Succinct Non-Interactive Argument of Knowledge (zkSNARKs) verification that is compiled as smart contracts to prevent malicious, compromised or even rational ICVs from uploading unreliable, erroneous or low-quality model updates. The verification process is embedded within the consensus of the underlying permissioned blockchain, which maximizes both the efficiency of the process and the utilization of computer resources. As demonstrated by discussions, security analysis, and numerical results, BV-ICVs reduced data poisoning attacks and increased the privacy protection and accuracy of FL.