Abstract
Federated Learning (FL) has emerged as a promising paradigm to address data silos and privacy concerns in artificial intelligence applications. It shows great potential in the Internet of Vehicles (IoV), where sensitive vehicle data must remain local. However, FL is highly vulnerable to model poisoning attacks, especially under non-IID distributions and high malicious participation. Existing defenses, most relying on gradient similarity across clients, often fail under such heterogeneous settings. To address this, we propose RDFDS, a robust defense framework for FL in IoV. RDFDS consists of two tightly coupled modules: (1) a fast distillation-based synthesis module that accelerates knowledge distillation via a lightweight preprocessing step, suitable for real-time vehicular environments; (2) a Gas-KMeans clustering aggregation module that adaptively assigns aggregation weights by identifying the benign majority through Gap Statistics, preserving reliable contributions without requiring global or cross-client reference data. Extensive experiments on MNIST, CIFAR-10, and GTSRB demonstrate that RDFDS consistently outperforms existing defenses in accuracy and robustness, particularly under extreme non-IID conditions and high malicious participation, highlighting its practicality for real-world IoV scenarios.
Funding Information
  • Guangzhou Science and Technology Planning Project (2024E04J1245)
  • Shanxi Key Laboratory of Industrial Internet Security, Taiyuan University of Technology (SKLIIS-2025-005)
  • National Natural Science Foundation of China (62472265)
  • National Natural Science Foundation of China (62502349)