A Tripartite Federated Learning Framework with Ternary Gradients and Differential Privacy for Secure IoV Data Sharing

The Internet of Vehicles (IoV) architecture generates vast amounts of distributed data that must be collaboratively learned while preserving privacy. This paper introduces a novel tripartite federated learning framework that significantly optimizes communication efficiency through ternary gradient quantization while enhancing privacy protection via federated differential privacy mechanisms. Our approach replaces traditional 32-bit floating-point gradient representations with efficient ternary values (-1,0,1), achieving over 93% reduction in communication overhead across multiple datasets. The proposed federated differential privacy scheme demonstrates superior performance compared to traditional centralized approaches, maintaining accuracy even under privacy-preserving constraints. Experimental evaluations across MNIST, Cifar10, Cifar100, and SVHN datasets validate the effectiveness of our approach, showing that tripartite gradients effectively reduce training gradient size by 93.33-93.74% while providing robust defense against member inference attacks. Comparative analysis with standard optimizers reveals that our method achieves faster convergence than SGD and higher stable-state accuracy than Adam, confirming that communication efficiency gains do not compromise model performance. This integration of tripartite gradient quantization with federated differential privacy offers a promising solution for secure, efficient collaborative learning in resource-constrained vehicular networks.