The Internet of Things (IoT) has been widely applied in various fields, and the free-space transmission of wireless links imposes security challenges for the identity authentication of IoT devices. Radio frequency fingerprint identification (RFFI) utilizes the subtle signal discrepancy caused by imperfections of wireless device hardware for identity authentication. However, the performance of RFFI is susceptible to variations in channels and receivers. To address this challenge, we adopt unsupervised contrastive learning and feature distribution alignment algorithms to tackle cross-domain feature shift. Experimental results show that the proposed scheme effectively leverages the feature distributions of unlabeled data from new receiver, achieving an average accuracy of 99.16%, which takes a step forward toward the practical application of RFFI.

Radio-Frequency Fingerprint Identification (RFFI) leverages unique hardware impairments of transmitters for physical-layer authentication. Despite its high security potential, deployed RFFI models suffer severe performance degradation when their receivers change, typically necessitating costly on-site fine-tuning or data re-collection. To alleviate this issue, this paper proposes a plug-and-play domain generalization (DG) framework that empowers existing RFFI systems with cross-receiver capabilities without retraining the backbone network. Specifically, a deployment-efficient receiver-agnostic extension (RAE) module is designed to augment the target RFFI system by cascading with its already trained models. To mitigate the impact of domain shifts, the RAE is trained using a novel adversarial feature alignment (AFA) strategy within a specialized model-agnostic meta-learning (MAML) framework. RAE aligns the feature distribution of unseen receivers with the source distribution expected by the trained model, thereby ensuring consistent generalization performance across different receivers. Experimental evaluations on real-world public datasets demonstrate that the proposed approach recovers approximately 31% of recognition accuracy (RA) losses and outperforms mainstream DG receiver-agnostic schemes by similar to 10%. Crucially, unlike existing methods that require architectural redesigns, our scheme offers a non-intrusive upgrade path for legacy RFFI systems.

ABSTRACT Physical‐layer‐specific emitter identification (SEI) is challenged by limited labelled data and the presence of previously unseen devices in practical deployments. To address these issues, we propose FS‐SA2SEI‐OSR, an open‐set extension of a few‐shot self‐supervised adversarial augmentation SEI framework. The proposed method combines adversarially trained few‐shot representation learning with continuous wavelet transform scalograms and lightweight open‐set recognition decision layers (OpenMax, EVM, OS‐SVM) to enable reliable classification of known emitters while explicitly rejecting unknown ones. Experiments on Wi‐Fi and public datasets (ADS‐B and FIT/CorteXlab) demonstrate a closed‐set accuracy of 89.9% and an open‐set detection rate of 64.0% at a worst‐case openness of 0.6, with only modest computational overhead. These results indicate that FS‐SA2SEI‐OSR provides a practical and scalable solution for few‐shot open‐set SEI in dynamic and resource‐constrained IoT environments.

Radio Frequency Fingerprint Identification (RFFI) has emerged as a promising secure authentication paradigm for Internet of Things (IoT) security. However, recent evidence suggests that deep-learning-based RFFI systems are critically vulnerable to adversarial attacks. Existing adversarial methodologies predominantly rely on white-box assumptions requiring full model transparency, which significantly limits their applications in real-world deployments. In this paper, we propose a novel Attention Reversal Attack (ARA) framework designed to generate highly transferable black-box adversarial examples. Our approach introduces a Jacobian-based Dataset Augmentation (JBDA) strategy to efficiently train high-fidelity surrogate models under a constrained query budget. Subsequently, we developed an ARA mechanism that misleads the model’s recognition results by suppressing its identification of key fingerprint features and redirecting attention to irrelevant background noise. Extensive experiments conducted on IoT dataset demonstrate that ARA achieves an average Attack Success Rate (ASR) > 90% across four architecturally diverse black-box models. These findings reveal a fundamental and pervasive security vulnerability in current RFFI designs, underscoring the urgent need for robust defense mechanisms in RFFI for physical-layer security.

We develop an active-passive algorithm using hard sample mining for label-free domain adapted radio-frequency fingerprint identification (RFFI). The average accuracy of RFFI is increased by ~7-14%, compared with the traditional baseline models.

For the physical-layer security enhancement in optical networks, hardware fingerprint authentication (HFA) is proposed by identifying subtle inherent impairments of different optical transmitters in the received signals using neural networks (NNs). However, due to NN's black-box nature, HFA may be susceptible to adversarial attacks, which is often overlooked in related research, but can lead to HFA erroneously granting high-level permissions to attackers. In this paper, we experimentally demonstrate that classical attacks against NNs, such as projected gradient descent (PGD), are also effective against HFA in optical networks, achieving an average targeted attack success rate of 66.43%. Furthermore, we propose an NN-based white-box attack framework named attack neural network (ATKNN) and a conditional generative adversarial network (CGAN)-based black-box attack framework named attack generator (ATKG). Both ATKNN and ATKG achieve a much higher success rate of similar to 100%, while their perturbations on the signals are significantly smaller and smoother, making the attacks more elusive. Our work reveals the shortcomings of the mainstream HFA frameworks and lays the foundation for HFA's future improvements.

Physical-layer secure key distribution (PLSKD) in fiber enhances security by utilizing channel dynamics to generate secure keys. However, PLSKD is significantly influenced by various channel scenarios. To improve the application of PLSKD, we investigate the use of transfer learning to enhance the performances. By leveraging the nonlinear mapping capabilities of neural networks using transfer learning, a high correlation coefficient and stable key generation rates across optical links of different fiber lengths are achieved, with lower computational requirements. Experimental results show that models pre-trained on a 100 km fiber link can be fine-tuned and effectively applied to a 300 km link, demonstrating excellent generalization and faster convergence with the same data volume. The proposed algorithm provides a practical solution for optimizing PLSKD performance in dynamic fiber links.

Radio frequency fingerprinting (RFF) has emerged as a lightweight solution for wireless device authentication by leveraging unique hardware-induced signal variations. We propose an efficient discriminative feature extraction scheme directly from raw Wi-Fi signals, based on a Siamese Convolutional Neural Network (CNN) deep learning framework. The model is trained on structured triplet data to optimize inter-class separation and intra-class compactness. Without relying on complex preprocessing techniques, the proposed approach achieves 95.06

To share the reciprocity of a single fiber channel in physical-layer secure key distribution (PLSKD), the bidirectional same wavelength (B-SW) transmission system is generally applied. However, the elastic Rayleigh backscattering noise severely degrades key performance after long fiber transmission. Here, the feasibility of a bidirectional wavelength division multiplexing (B-WDM) transmission in a single fiber for high-speed long-range PLSKD is verified, where the wavelength dependence of fiber channel reciprocity is evaluated. PLSKD based on B-WDM in one standard 100 GHz WDM channel is demonstrated, offering a key generation rate of 4.8 Gbps over a 200 km fiber link, providing a 200% increment in terms of key generation rate (KGR) compared with B-SW transmission.

Radio frequency fingerprint identification (RFFI), which leverages hardware-specific impairments in Internet of Things (IoT) devices, is widely used for device authentication and spoofing attack detection to enhance communication security. However, the existing RFFI methods heavily depend on large-scale training datasets in deep learning (DL), with severe overfitting issues if the training samples are scarce. This article proposes a few-shot learning framework that combines feature augmentation and meta-learning to overcome these challenges. A novel data augmentation technique based on grad class activation maps (Grad-CAM) is introduced to address the scarcity of training samples, which generates augmented samples by adjusting the weights of receptive fields in feature maps, forming an auxiliary dataset for training. In meta-training, the auxiliary dataset is used to construct tasks comprising support and query sets. By extracting common features from the limited sample size, the framework trains a meta-model with robust generalization capabilities. In the deployment phase, a fine-tuning strategy further optimizes the classifier using a small labeled dataset from new IoT devices, allowing rapid adaptation with high accuracy. The proposed framework is evaluated on the large-scale open-source dataset, achieving an accuracy of 94.1% under 8-way 5-shot, with only 25 samples per device in meta-training. Meta-learning boosts performance by 15%-20%, with meta-training feature augmentation further increasing accuracy by 5%-6.6%. Compared to baseline methods, the proposed framework improves the accuracy by 30%, which outperforms the state-of-the-art algorithms by 4%-25%.

Radio frequency fingerprint (RFF) authentication is promising for physical-layer hardware identification. However, RFF identities exposed to public channels are easily susceptible to fake identification attacks. To render the unique RFF features inaccessible to unauthorized devices, we propose an active security approach based on chaotic fingerprints, which can generate 10320 chaotic fingerprints against the active identification attacks using fake RFF hardware identities. The proposed scheme is experimentally demonstrated, dramatically reducing the recognition accuracy (RA) to <11% for the unauthorized users, while the success rate (SR) of RFF forgery attacks is reduced to <9%. Instead, an RA >90% at 20dB is feasible for the authorized devices, which implies the successful recovery of the original RFF. The proposed chaotic RFF significantly enhances the high security level of identity authentication for wireless communications.

Physical-layer secure key distribution (PLSKD) exploiting channel reciprocity and randomness is promising for data security enhancement in fiber, offering an endogenous high key generation rate (KGR) up to Gb/s. The PLSKD provides the integration possibility of key generation and data transmission within the same fiber channel. The coexistence schemes of PLSKD and secure data transmission in fibers are proposed using wavelength division multiplexing (WDM), and full-duplex transmission at the same wavelength. The coexistence schemes are experimentally demonstrated with a 3.4 and 3.25 Gb/s PLSKD, where the encrypted data rate is 15.7 Gb/s, over a 20 km standard single-mode fiber (SSMF) using WDM or full-duplex respectively. The effects of channel spacing using WDM and trade-offs between PLSKD and secure data transmission using full-duplex are fully analyzed. The proposed secure data transmission with PLSKD is classical and fully compatible with current fiber infrastructures.

Neural networks are applied for fiber mapping in unidirectional physical-layer key distribution, where the bi-directional long short-term memory increases the key generation rate from 12.28 to 13.68 Gbps.

Robust radio frequency fingerprint identification (RFFI) is crucial for physical layer authentication, while it suffers from channel effects and requires extra overhead to increase recognition accuracy (RA). To address this, an efficient channel-robust RFFI scheme is proposed, employing a specialized multi-task learning (MTL) framework to direct the neural network (NN) toward extracting channel-robust features. In addition, receiver collaboration (RC) is utilized for data augmentation and output calibration. Experimental results demonstrate that the RA is significantly increased from 51.72% to 99.97% when using the open-resource Wi-Fi signal datasets collected from different time periods. Meanwhile, the requirements for extra data transmission, NN structure, and feature crafting in the inferring stage are dramatically simplified.

A chaotic discrete Fresnel transform (DFnT) precoding scheme is proposed for physical-layer security enhancement in orthogonal frequency division multiplexing passive optical networks (OFDM-PONs). A multi-fold OFDM data encryption is applied in terms of the chaotic scale factor, and chaotic row/column permutations in the DFnT matrix, and the chaos sensitivity creates a huge key space of 10138. The peak-to-average power ratio (PAPR) performances of OFDM signals associated with a scale factor in DFnT are fully analyzed. A 38.2 Gb/s OFDM signal transmission with chaotic DFnT is successfully demonstrated over a 20 km standard single-mode fiber (SSMF), where the gain of receiver sensitivity is ∼2 dB (BER @10−3), compared with OFDM without precoding.

Physical-layer secure key distribution (PLSKD) generally acquires highly correlated entropy sources via bidirectional transmission to share the channel reciprocity. For long-haul fiber links, the non-negligible backscattering noise (BSN) and the challenge of bidirectional optical amplification degrade the key generation performances. Since the channel reciprocity can be precisely mapped using neural networks (NNs), unidirectional PLSKD provides a feasible PLSKD for longer fiber links. Here, a final error-free key generation rate (KGR) in unidirectional PLSKD of 3.07 Gb/s is demonstrated over a 300 km fiber link using NNs. Moreover, the channel mapping is analyzed in terms of fiber distance, chromatic dispersion, the nonlinearity of random source, and BSN.

A security approach is proposed to mask the specific emitter identification (ID) attacks by generating 10 320 chaotic ID features. The success rate of forgery attacks is <11%, and the recognition accuracy >90% at 20 dB is verified for the authorized devices.

Physical-layer authentication (PLA) based on hardware fingerprints can safeguard optical networks against large-scale masquerade or active injection attacks. However, traditional schemes rely on massive labeled close-set data. Here, we propose an unsupervised hardware fingerprint authentication based on a variational autoencoder (VAE). Specifically, the triplets are generated through variational inference on unlabeled optical spectra and then applied to train the feature extractor, which has an excellent generalization ability and enables fingerprint feature extraction from previously unknown optical transmitters. The feasibility of the proposed scheme is experimentally verified by the successful classification of eight optical transmitters after a 20 km standard single-mode fiber (SSMF) transmission, to distinguish efficiently the rogue from legal devices. A recognition accuracy of 99% and a miss alarm rate of 0% are achieved even under the interference of multiple rogue devices. Moreover, the proposed scheme is verified to have a comparable performance with the results obtained from supervised learning.