Cross-Receiver Generalization for RF Fingerprint Identification via Feature Disentanglement and Adversarial Training

TL;DR

This work tackles cross-receiver generalization in RF fingerprint identification (RFFI) by proposing DRIFT, a disentangled representation framework that separates transmitter-specific from receiver-specific features and uses adversarial training (via a gradient reversal layer) plus a style-transfer–inspired regularization to purify receiver features. By enforcing domain-invariant transmitter representations while stabilizing receiver characteristics, DRIFT achieves robust performance across unseen receivers and across days, with substantial gains over state-of-the-art baselines. Theoretical analysis based on HDivergence and convex-hull formulations provides a justification for reducing domain divergences, and extensive experiments on the WiSiG ManySig dataset demonstrate high adaptability to new receivers and temporal variation. The method has practical impact for scalable, maintenance-friendly RFFI deployments in IoT and wireless security, reducing the need for retraining when hardware changes occur.

Abstract

Radio frequency fingerprint identification (RFFI) is a critical technique for wireless network security, leveraging intrinsic hardware-level imperfections introduced during device manufacturing to enable precise transmitter identification. While deep neural networks have shown remarkable capability in extracting discriminative features, their real-world deployment is hindered by receiver-induced variability. In practice, RF fingerprint signals comprise transmitter-specific features as well as channel distortions and receiver-induced biases. Although channel equalization can mitigate channel noise, receiver-induced feature shifts remain largely unaddressed, causing the RFFI models to overfit to receiver-specific patterns. This limitation is particularly problematic when training and evaluation share the same receiver, as replacing the receiver in deployment can cause substantial performance degradation. To tackle this challenge, we propose an RFFI framework robust to cross-receiver variability, integrating adversarial training and style transfer to explicitly disentangle transmitter and receiver features. By enforcing domain-invariant representation learning, our method isolates genuine hardware signatures from receiver artifacts, ensuring robustness against receiver changes. Extensive experiments on multi-receiver datasets demonstrate that our approach consistently outperforms state-of-the-art baselines, achieving up to a 10% improvement in average accuracy across diverse receiver settings.

Figures (3)

• Figure 1: Illustration of transmitter and receiver hardware impairments.
• Figure 2: Overview of the system model.
• Figure 3: Sensitivity analysis of hyperparameters $\lambda_1$, $\lambda_2$, and $\lambda_3$.

Theorems & Definitions (1)

• Theorem 1: Generalization Bound on Unseen Domain Risk