Robust decentralized federated learning and its application in intelligent transportation systems
Date
2025
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
Federated learning (FL) has become a foundational paradigm for privacy-preserving, collaborative machine learning across decentralized data sources. By enabling distributed devices to train models without sharing raw data, FL addresses critical privacy and regulatory concerns in domains such as healthcare, finance, edge computing, and the Internet of Things. However, practical deployment of FL faces two fundamental challenges: ensuring robustness and trust in adversarial or untrusted environments, and extending FL to real-world, large-scale domains such as intelligent transportation systems, which generate vast, distributed, and dynamic data. ☐ This dissertation, titled Robust Decentralized Federated Learning and Its Application in Intelligent Transportation Systems, is structured around two complementary research directions that collectively advance the state of the art in decentralized FL. The first research direction focuses on blockchain-based robust federated learning. In this part, I develop and analyze two frameworks, VBFL and NBFL, that leverage blockchain technology to address the challenges of trust, security, and robustness in collaborative learning. Traditional FL frameworks often rely on a central coordinator and lack model validation mechanisms, which introduces a single point of failure and makes the system vulnerable to adversarial attacks such as model poisoning and free-riding. To overcome these limitations, VBFL and NBFL eliminate central coordination by employing decentralized consensus protocols based on blockchain, such as proof-of-stake and validator voting. These protocols probabilistically favor the aggregation of legitimate, high-quality model updates while reducing the influence of malicious or low-quality contributions. Both frameworks incorporate robust validation and incentive mechanisms that reward honest participation and penalize adversarial behaviors. NBFL also enables personalized and communication-efficient learning through structured model pruning inspired by the Lottery Ticket Hypothesis, while demonstrating enhanced generalization and robustness in Non-IID environments compared to baselines. The effectiveness and accountability of VBFL and NBFL are demonstrated through extensive experiments on benchmark datasets, showing resilience and adaptability in adversarial and heterogeneous environments. ☐ The second research direction investigates decentralized federated learning in intelligent transportation systems (ITS). In this context, I propose and evaluate two novel FL frameworks, BFRT and NeighborFL, that enable real-time, adaptive, and collaborative learning among distributed edge devices in large-scale ITS networks. BFRT is designed for blockchain-compatible federated real-time traffic prediction, supporting online model updates and multi-step forecasting using streaming traffic data. NeighborFL introduces individualized aggregation strategies, allowing each device to form dynamic, context-aware groups of collaborators based on spatial proximity and real-time prediction error reduction. These approaches enable improved local traffic prediction performance, and support real-time streaming data scenarios with online model updates. NeighborFL facilitates adaptive neighbor selection, which is crucial for responsive and robust traffic prediction. The effectiveness and scalability of BFRT and NeighborFL are validated through real-world ITS datasets and simulations, demonstrating their strength over centralized and naive federated baselines in terms of prediction accuracy. ☐ These two research directions converge on the goal of building robust, scalable, and trustworthy federated learning in decentralized settings. The blockchain-based frameworks, VBFL and NBFL, ensure trust, auditability, and resilience in adversarial environments. Meanwhile, the decentralized, blockchain-compatible FL protocols, BFRT and NeighborFL, tackle real-time challenges in large-scale ITS. Together, they establish a solid foundation for secure, efficient, and generalizable federated learning, with applications extending beyond transportation to domains like smart cities, energy systems, and personalized education.
Description
Keywords
Intelligent transportation systems, Federated learning, Blockchain technology