Information Bottleneck-Driven Cross-Receiver RF Fingerprinting for Physical-Layer Security

Radio frequency fingerprinting (RFF) is a promising technique for physical-layer security. It exploits hardware-induced distortions as fingerprint features to enable passive, keyless, and protocol-agnostic device authentication. Recent advances in RFF have significantly enhanced its performance and efficiency. Nevertheless, the cross-receiver scenario remains a persistent challenge because heterogeneous receivers introduce complex and nonlinear distribution shifts that hinder model generalization. Existing methods primarily attempt to address this challenge through domain adaptation (DA) or domain generalization (DG), which typically assume access to target-receiver data or receiver labels that are often difficult or expensive to obtain in practice. Moreover, these methods are generally grounded in an idealized assumption of domain invariance, under which the feature distributions derived from signal samples collected by different receivers are expected to be aligned in a shared latent space. However, this assumption may be violated in practice due to the complex and nonlinear discrepancies across receivers, which can result in over-alignment and consequently degrade the discriminative capability of RFF models. To address these challenges, we propose an information-theoretic framework for cross-receiver RFF that departs from domain-invariant assumptions. Within the framework, a lightweight frequency-aware network (FAN) serves as the backbone, integrating frequency-aware adaptive filtering and dual-stream convolutional modules to balance representational capacity and computational efficiency. More importantly, the information bottleneck (IB) principle is employed to learn compact and discriminative representations while mitigating receiver-induced redundancy, without relying on receiver labels. Extensive experiments on the WiSig dataset demonstrate the effectiveness of the proposed method, which consistently outperforms existing methods under both cross-receiver and cross-receiver-and-day settings, and is further validated on the LoRa dataset under the joint cross-day and cross-receiver setting. The source code can be available at https://github.com/BeechburgPieStar/IB-RFF