Co-Evolutionary Flow Matching for Cross-Domain RF Fingerprint Calibration: A Plug-and-Play Tool

Abstract

The domain shift caused by dynamic wireless environments is one of the fundamental bottlenecks for cross-domain radio frequency fingerprint identification (RFFI). Traditional adversarial disentanglement methods avail of orthogonality between RF fingerprint (FP) and wireless environments, yet they ignore the non-orthogonal physically-coupled nature between them, which leads to loss of physical continuity during alignment and easily triggers negative transfer. This letter proposes a physics-inspired Co-Evolutionary Flow Matching (Co-FM) architecture, which redefines cross-domain RFFI as a continuous manifold transformation guided by wireless environments in latent space. Specifically, we use residual physical correlations from incomplete disentanglement as evolutionary driving force, and define Domain Identifier (DI) features as exogenous energy guiding operators to steer the evolution trajectory of FP feature in unsupervised target domains. Co-FM is a lightweight plug-and-play tool, which can be directly integrated into the existing pre-trained disentanglement models without retraining. Experiments on the self-built DRFF-R2 dataset demonstrate that Co-FM achieves a maximum overall accuracy improvement of 8.05% under the MobileNetV2 backbone. Concurrently, under the ResNet backbone, where adversarial alignment encounters negative transfer, Co-FM eliminates this degradation and boosts the weak-category identification rate from 0% to 14.84%, while maintaining an ultra-low latency increment of less than 0.01 ms.