Channel-Robust Specific Emitter Identification via Complex-Valued Contrastive Transformers

Specific Emitter Identification (SEI) authenticates a transmitter from the hardware-intrinsic fingerprint it leaves on its signal, which makes it attractive for physical-layer security. The difficulty is that these analog impairments are faint, and in high-mobility links they must be read through severe time-selective Rayleigh fading with Doppler. Real-valued deep networks make this harder than it needs to be: splitting the in-phase and quadrature branches weakens the joint phase-amplitude coupling, and cross-entropy training leaves decision boundaries that drift as the channel changes. This paper takes a different route with a Complex-Valued Transformer (CV-Trans) trained under a Supervised Complex Contrastive Learning (SupCon) objective. The model stays in the complex domain throughout: it fuses the raw envelope with differential-trace phase features to retain the In-phase/Quadrature (I/Q) mismatch and phase-noise signatures, and the SupCon loss tightens each emitter's cluster on a unit hypersphere so that the fingerprint separates from stochastic Doppler rotation. Under Rayleigh fading with Doppler, the framework reaches 99.97% accuracy at 20 dB SNR and holds 99.7% at high mobility (fdTs=10−2), while using 50% fewer trainable linear parameters than the real-valued equivalent and running in about 2.73 ms per burst – fast enough for real-time use.