Interpretable reputation driven asynchronous consensus vehicle networking federated learning architecture.

Area of Science:

• * Vehicular communication networks
• * Distributed artificial intelligence
• * Cybersecurity

Background:

• * Federated learning in the Internet of Vehicles (IoV) faces challenges including lack of interpretability, unreliable reputation evaluation, and centralized architecture vulnerabilities.
• * Existing systems struggle with Differential Privacy (DP), Non-Independent and Identically Distributed (Non-IID) data, and Byzantine attacks.
• * There is a need for a robust and interpretable framework for collaborative training in high-concurrency vehicular environments.

Purpose of the Study:

• * To propose a Vehicle-Road-Cloud-Chain (VRCC) four-layer collaborative framework for federated learning in IoV.
• * To enhance interpretability, reputation evaluation, and address centralization vulnerabilities.
• * To ensure robustness against DP, Non-IID data, and Byzantine attacks while supporting high concurrency.

Main Methods:

• * Developed an asynchronous blockchain consensus layer using Directed Acyclic Graph (DAG) for low-latency model interaction storage.
• * Designed a three-layer interpretable reputation evaluation mechanism incorporating historical performance, MMD Bayesian inference, and contribution.
• * Proposed a participant selection algorithm using Deep Deterministic Policy Gradient (DDPG) and reputation partitioning for optimized resource allocation and model quality verification.

Main Results:

• * The proposed VRCC framework demonstrates explicit decoupling between data sharing and collaborative training.
• * Reputation judgments are made auditable and transparent through joint signing and on-chain uploading.
• * The participant selection algorithm effectively optimizes communication overhead, computation delay, and redundant filtering.
• * Experimental results show rapid convergence and stable cumulative rewards, indicating system-level superiority.

Conclusions:

• * The VRCC framework provides an interpretable and robust solution for federated learning in IoV.
• * The system achieves high-concurrency scalability and superior performance in aggregation and verification.
• * The proposed methods effectively address the challenges of DP, Non-IID data, and Byzantine attacks in vehicular networks.