On-Device Multimodal Federated Learning for Efficient Jamming Detection

Abstract: Wireless networks face severe vulnerabilities from jamming attacks, which can significantly disrupt communication. Existing detection approaches are often unimodal, rely on centralized processing, and demand substantial computational resources, hindering scalability, efficiency, and deployment feasibility. To address these challenges, we introduce a multimodal Federated Learning (FL) framework for on-device jamming detection and classification that integrates spectrograms with cross-layer network Key Performance Indicators (KPIs) through a lightweight dual-encoder architecture equipped with a fusion module and a multimodal projection head. This design enables privacy-preserving training and inference by ensuring that only model parameters are exchanged, while raw data remains on the device. The framework is implemented and evaluated on a wireless experimental testbed using, to the best of our knowledge, the first over-the-air multimodal dataset with synchronized benign and three distinct jamming scenarios. Results show that our approach surpasses state-of-the-art unimodal baselines by up to 15% in detection accuracy, achieves convergence with 60% fewer communication rounds, and maintains low resource usage. Its benefits are most evident under heterogeneous data distributions across devices, where it exhibits strong robustness and reliability.