Exploiting Radio Frequency Fingerprints for Device Identification: Tackling Cross-receiver Challenges in the Source-data-free Scenario

TL;DR

This work addresses the practical challenge of cross-receiver RF fingerprint identification in edge scenarios when source-domain data cannot be shared. It introduces SCRFFI and proposes MS-SHOT, a source-data-free adaptation framework that refines pseudo-labels with momentum-center soft labeling and enforces global class-proportion constraints to combat label shift and imbalanced targets. The authors establish generalization bounds showing target performance hinges on pseudo-label accuracy and demonstrate, via real-worldWisig and HackRF experiments, that MS-SHOT consistently outperforms existing methods and remains robust under noise, scale, and unknown priors. The results highlight MS-SHOT as a scalable, practical solution for secure device authentication in heterogeneous, privacy-preserving edge environments.

Abstract

With the rapid proliferation of edge computing, Radio Frequency Fingerprint Identification (RFFI) has become increasingly important for secure device authentication. However, practical deployment of deep learning-based RFFI models is hindered by a critical challenge: their performance often degrades significantly when applied across receivers with different hardware characteristics due to distribution shifts introduced by receiver variation. To address this, we investigate the source-data-free cross-receiver RFFI (SCRFFI) problem, where a model pretrained on labeled signals from a source receiver must adapt to unlabeled signals from a target receiver, without access to any source-domain data during adaptation. We first formulate a novel constrained pseudo-labeling-based SCRFFI adaptation framework, and provide a theoretical analysis of its generalization performance. Our analysis highlights a key insight: the target-domain performance is highly sensitive to the quality of the pseudo-labels generated during adaptation. Motivated by this, we propose Momentum Soft pseudo-label Source Hypothesis Transfer (MS-SHOT), a new method for SCRFFI that incorporates momentum-center-guided soft pseudo-labeling and enforces global structural constraints to encourage confident and diverse predictions. Notably, MS-SHOT effectively addresses scenarios involving label shift or unknown, non-uniform class distributions in the target domain -- a significant limitation of prior methods. Extensive experiments on real-world datasets demonstrate that MS-SHOT consistently outperforms existing approaches in both accuracy and robustness, offering a practical and scalable solution for source-data-free cross-receiver adaptation in RFFI.

Figures (8)

• Figure 1: An RFFI-based access authentication system workflow. The diagram shows the workflow of an RFFI-based access authentication system. IoT devices (e.g., smartphones, cameras, drones) transmit wireless signals, which are captured by an authentication access point (receiver). The system performs signal processing to filter out noise and enhance signal quality, followed by feature extraction and classification to identify unique RF fingerprints. Based on the classification results, the system decides whether to grant or deny access to the device.
• Figure 2: The scenario of source-data-free cross-receiver RFFI. Prior training commences in the source domain (Rx-1), where the RFFI model is developed using labeled signal data. Following model training in the source domain, a transfer phase occurs where only the learned model is applied to the target domain (Rx-2) due to data privacy. In the target domain, the model undergoes adaptation using signals without associated labels, resulting in an adapted RFFI model adapted to the new receiver.
• Figure 3: The overview of MS-SHOT. The parameters of the adapted model $h^t$ are first initialized with the feature extractor parameters $\theta_{\sf F}^s$ and classifier parameters $\theta_{\sf C}^s$ of the model trained in the source domain. Following this, the classifier parameters in $h^t$ are frozen to preserve the decision boundary, while updates are confined to the feature extractor parameters. Throughout the adaptation phase, the model interacts with Momentum Center guided Soft Pseudo-labeling (MCSP) during each batch to procure real-time soft pseudo-labels $\tilde{y}^{\rm soft}$. Loss functions $\mathcal{L}_{ce}$, $\mathcal{L}_{nn}$, and $\mathcal{L}_{\ell_1}$ are then employed to calculate the loss, and gradient descent is utilized to iteratively update $\theta_{\sf F}^t$ until the model achieves convergence (FC: Fully-Connected layer, BN: BatchNorm, CLS: Classifier).
• Figure 4: Scenario description of Wisig dataset hanna2022wisig. (a) Nodes are arranged as a grid. Each node is a roof-mounted PC with at least one WiFi emitter (Tx). Some nodes are equipped with USRP receivers (Rx). (b) Positions of Tx. (c) Positions of Rx.
• Figure 5: Experimental results of different hyper-parameters on model performance. (a) Weight of $\mathcal{L}_{ce}$ ($\lambda_1$). (b) Weight of $\mathcal{L}_{nn}$ ($\lambda_2$). (c) Weight of $\mathcal{L}_{\ell_1}$ ($\lambda_3$). (d) Momentum of the feature center update ($\beta$)

Theorems & Definitions (5)

• Theorem 1
• Theorem 2
• Corollary 1
• Theorem 3
• Remark 1