Robust Radio Frequency Fingerprint Identification via Multi-Scale Complex-Valued Learning and Balanced Domain Alignment

Abstract

Specific emitter identification (SEI) enables physical-layer authentication of wireless devices by exploiting hardware-induced radio-frequency (RF) fingerprints embedded in received signals. However, existing deep learning–based SEI methods often suffer from limited cross-domain generalization, as models tend to overfit channel- and receiver-dependent characteristics rather than stable hardware-related fingerprints. To address this challenge, this paper proposes a robust SEI framework that integrates a Multi-Scale Complex-Valued Neural Network (MSCVNN) with a Balanced Maximum Mean Discrepancy (BMMD)–based domain alignment strategy. Specifically, we employ two complementary branches: one operates on the raw signal to preserve amplitude–phase waveform cues, and the other uses empirical mode decomposition (EMD) to obtain multi-scale time–frequency representations. The extracted complex features are then modeled by the MSCVNN through multi-scale complex-valued convolutions and a multilayer perceptron, capturing both local structures and global dependencies in the complex domain. In addition, BMMD aligns class-conditional feature distributions across multiple receiver domains, alleviating domain shift without relying on target-domain data and promoting discriminative, domain-invariant representations. Extensive experiments conducted on a selected WiSig subset, where twelve USRP receivers are organized into six receiver domains with two receivers in each domain, demonstrate that the proposed method improves cross-receiver identification performance compared with representative backbone networks and domain-alignment baselines. In addition, the proposed framework exhibits strong robustness under low signal-to-noise ratio and feature-masking conditions, validating its effectiveness in challenging and dynamic wireless environments.