Open-Set RF Fingerprint Recognition via Conditional Variational Adversarial Learning with Complex-Valued Networks

To address the security threats of device identity spoofing and unauthorized access in wireless communications, machine-learning-based radio frequency fingerprint (RFF) technology provides a promising physical-layer solution for device authentication. However, most existing deep learning RFF methods rely on a closed-set assumption, limiting their applicability in practical open-set Internet of Things (IoT) scenarios involving unregistered transmitters. This paper proposes a robust openset RFF identification framework based on a Conditional Variational Autoencoder-Generative Adversarial Network (CVAE-GAN). Specifically, we integrate an attention mechanism with a complex-valued convolutional network to natively preserve the intrinsic phase-amplitude coupling of I/Q signals, thereby extracting highly robust, device-specific features under noisy conditions. Furthermore, by explicitly introducing variational latent regularization, the CVAE-based generator effectively mitigates mode collapse and synthesizes diverse, high-fidelity boundary samples to significantly tighten the discriminator’s open-set decision boundaries. Extensive experiments using commercialWi- Fi devices demonstrate that the proposed method achieves state-of-the-art closed-set accuracy. In open-set scenarios, it maintains robust unknown rejection capabilities even with an additional 5 dB reduction in signal-to-noise ratio (SNR) compared to baseline methods. Additionally, the generative module successfully deceives independently trained classifiers with an average success rate of 92.9%, validating its superior synthesis fidelity and essential contribution to the overall adversarial framework.