Advancing Federated Learning: A Comprehensive Solution for Model Aggregation, Heterogeneity, Privacy, and Security

Secure vehicular communication is one of the challenges that is crucial for Intelligent Transportation Systems (ITS). Although Federated Learning (FL) enhances data privacy compared to Centralized Learning (CL), it faces challenges with data heterogeneity and resource allocation. To address these challenges, the Hybrid Federated-Centralized Learning (HFCL) framework improves efficiency and reduces communication overhead but still lacks comprehensive solutions to existing key issues including FL model aggregation, client heterogeneity, privacy-preserving communication, and enhanced security protocols. This paper proposes a unified HFCL framework incorporating the existing key issues that aims to develop and analyze model aggregation ensuring equitable client contributions, secure communications, and protect against various security threats. The implementation with FedAvg aggregation algorithm on diverse clients and datasets ensures robust and efficient communication, while Differential Privacy (DP) in HFCL enhances security and resilience against adversaries including poisoning attacks. Performance analysis reveals that HFCL maintains high accuracy and stability across varying learning rates and client configurations. Notably, the proposed unified privacy preserving HFCL achieves nearly 99.15% accuracy on the MNIST dataset and 68.4% on the CIFAR10 dataset, demonstrating its efficacy in different scenarios. By addressing the existing critical challenges, the proposed framework enhances the scalability, performance, and security of distributed ML systems, with a primary focus on vehicular communication. It highlights the significant advantages of HFCL for data processing and decision-making in applications like collision avoidance and autonomous driving. Additionally, the framework’s applicability could be extended to healthcare and finance, enabling secure and efficient analysis of sensitive data. © The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd. 2025.