Receiver-Agnostic Radio Frequency Fingerprint Identification via Federated Learning

Ensuring secure and reliable wireless connectivity is essential for modern Internet of Things (IoT) applications. Radio frequency fingerprint identification (RFFI) has emerged as a promising lightweight device authentication mechanism by leveraging unique hardware-induced features in transmitted signals. This paper proposes a federated RFFI framework specifically designed to tackle open challenges associated with receiver drift, label skewed data distribution and client selection. The framework introduces a receiver-agnostic training scheme based on adversarial learning in a distributed setting, enabling the global model to suppress receiver-specific features while retaining transmitter-distinctive representations. Evaluations on a real-world dataset confirm that the proposed federated RFFI framework achieves improved transmitter classification accuracy on previously unseen receivers by up to 40% compared to baseline non-adversarial approach. It also presents a systematic analysis of label-skewed data distributions, revealing that model performance degrades as skew increases and motivating the development of strategies to address this issue. To that end, a Label Loss Driven client selection strategy is proposed, which prioritizes the most informative clients based on their contribution to transmitter classification accuracy, resulting in faster convergence and improved generalization. Under high label skew, the proposed client selection strategy achieves a convergence improvement of 49–51% over baselines, with communication overhead reduced by 27–49% and computation overhead by about 50%. Overall, this work provides a practical and effective solution for deploying RFFI in scalable, resource-constrained IoT systems.