Feature Alignment Transformer for Cross-domain Radio Frequency Fingerprint Identification

Radio Frequency Fingerprint Identification (RFFI) has emerged as an effective device authentication paradigm, offering non-cryptographic, tamper-proof, and lightweight solutions for IoT networks. However, conventional deep learning-based RFFI schemes often experience significant performance degradation in cross-domain scenarios due to the distribution shift between training and test datasets, known as the cross-domain RFFI problem. To address this issue, we propose FATransformer, a novel framework for cross-domain RFFI, which comprises four key modules: data preprocessing, transformer backbone, feature alignment, and automatic tuning. Additionally, we establish a theoretical relationship between target domain error bounds and domain discrepancies to support the design of the feature alignment method. Extensive experiments on real-world datasets demonstrate FATransformer’s superior accuracy in multiple cross-domain RFFI tasks compared to baselines.