Interpretable reputation driven asynchronous consensus vehicle networking federated learning architecture.

This paper proposes a Vehicle-Road-Cloud-Chain (VRCC) four-layer collaborative framework to address the issues of lack of interpretability, reputation evaluation failure, and architecture centralization vulnerability faced by federated learning in Internet of Vehicles (IoV) under Differential Privacy (DP), Non-Independent and Identically Distributed (Non-IID) data, and Byzantine attacks. The framework achieves explicit decoupling between low-latency data sharing and latency-tolerant collaborative training. Firstly, construct an asynchronous blockchain consensus layer based on Directed Acyclic Graph (DAG), which supports low confirmation latency model interaction record storage in high-concurrency vehicle scenarios; And design a three-layer interpretable reputation evaluation mechanism, integrating historical task performance, Maximum Mean Discrepancy (MMD) Bayesian inference, and task completion contribution, to achieve causal decoupling between "honest high loss" and "malicious low loss reporting" under Differential Privacy noise, and jointly sign and upload it to the chain through the regulatory committee and Roadside Units (RSUs), making reputation judgments auditable and transparent; Further propose a participant selection algorithm based on Deep Deterministic Policy Gradient (DDPG) and reputation partitioning, which synchronously optimizes communication overhead, computation delay, and redundant filtering in dynamic traffic flow, while utilizing local DAG weight-biased random walks to achieve lightweight asynchronous model quality verification. The experiment shows that the cumulative reward of the proposed method can quickly converge and remain stable, verifying its system-level superiority in interpretable robust aggregation and high-concurrency scalability.