Robust Multi-Modal RF Fingerprinting in Intelligent Reality Using Federated Learning

Radio frequency (RF) fingerprinting enables the identification of wireless transmitters by exploiting hardware-induced signal distortions unique to each device, a critical capability for secure device recognition in Intelligent Reality environments such as augmented and virtual reality systems, while federated learning (FL) has emerged as a privacy-preserving alternative to centralized training for RF fingerprinting, allowing access points (APs) to collaboratively learn models without sharing raw signal data. However, existing FL methods typically rely on in-phase and quadrature (IQ) time-domain samples alone, which limits performance, slows convergence, and increases communication overhead, especially in non-i.i.d. environments prevalent in dynamic Intelligent Reality deployments. In this work, we propose FUSE-RF, a robust federated learning framework that incorporates multi-modal signal representations, including time, frequency, and amplitude-phase domains, to enhance RF fingerprinting for intelligent reality applications, leveraging complementary features across modalities to reduce local gradient variance, improve model generalization, and enable real-time, scalable device authentication in immersive IoT ecosystems. We theoretically demonstrate improved convergence behavior under this multi-modal setup and empirically validate our framework on a large-scale real-world RF dataset, with results showing that FUSE-RF achieves up to 30% higher accuracy and requires significantly fewer communication rounds compared to single-modal FL baselines, making it an efficient solution for distributed RF device identification in Intelligent Reality.