With the deployment of 5G and large-scale Internet of Things (IoT), the equipment identification and authentication scheme based on RF fingerprint shows unique advantages in terms of lightweight and uniqueness. However, traditional RF fingerprint identification scheme based on machine learning has the disadvantages of high computational complexity and low accuracy. Meanwhile, this scheme requires large-scale labeled datasets to realize network learning, and due to the nonlinearity of the cascade, we can not well understand the properties and optimal configurations of these networks. To solve above problems, in this paper, we propose an RF fingerprint identification method based on wavelet scattering network in the small-scale dataset. Specifically, in this method, we first design a hybrid network model of wavelet scattering network combined with deep residual network (Resnet18). Then, since one of the main problems of RF fingerprinting is the diversity of signal information at different time scales, we choose to use the construction of scattering network based on wavelet basis to complete the accurate feature decomposition of the nonlinear features of RF fingerprint. These features are stable against deformations and retain high frequency information for identification. Finally, we can use the obtained detailed features to realize the accurate identification of RF radiation source equipments. The experimental results show that our scheme can better suppress the interference of noise in the signal, improve the feature representation ability, and it can obtain higher identification accuracy than other comparison schemes.

Deep learning (DL) based identification systems are deemed as the scalable, accurate and lightweight authentication mechanisms to handle the security provisioning of massive Internet of Things (IoT) systems by leveraging the hardware-level radio frequency fingerprints. However, the conventional DL-based methods perform poor generalization in the practical time-varying signal-to-noise ratio (SNR) scenarios. In this paper, we propose a data augmentation and contrastive learning based radio frequency fingerprinting (DACL-RFF) with the joint optimization of samples agreement and labels agreement. First, we expand the SNR variations of training dataset with data augmentation, and then we propose a novel framework of contrastive learning. Specifically, we employ the original samples as the supervisory information of augmented samples and the label information of original samples is leveraged to guide the training process. Experimental results demonstrate that our proposal can increase the average accuracy by up to 51.74% in comparison with the case of none augmentation as the conventional DL-based methods. Additionally, we show that our framework of contrastive learning yields 5.27% improvement compared to the case of data augmentation with supervised learning.

Symbiotic communications have been deemed as a critical technology for Internet of Things (IoT) communications owing to its high spectrum and energy efficiency. Reconfigurable intelligent surface (RIS), which can tune wireless transmission channels by manipulating incident waves through the corresponding electromagnetic elements, is a promising enabler of various symbiotic communications scenarios in IoT. However, when the full electromagnetic elements of RIS are activated, system capacity will be improved and energy efficiency will be reduced inevitably, also with undesirable power consumption. To address this issue, an intelligent dynamic subarray RIS framework based on deep reinforcement learning (DRL) has been proposed. The key idea is to divide RIS electromagnetic elements into several groups and optimize power amplifier factor, independent phase shifts to improve the system energy efficiency under the premise of user’s basic requirements. In particular, we formulate a hybrid optimization problem of RIS subarray partition and beamforming to maximize system energy efficiency. It can be proved that this hybrid optimization is a mixed nonconvex integer programming problem. To solve this issue, we proposed a comprehensive DRL framework, including two parts, i.e., 1) a Markov decision process (MDP) to model the subarray partition design, amplitude, and phase shifts of RIS and 2) an active RIS subarray optimization scheme based on deep deterministic policy gradient. Numerical results have demonstrated that, compared with the conventional fully-connected RIS, the system energy efficiency can be significantly improved.

Radio frequency fingerprints (RFF) identification is a critical enabling technology to support rapid and scalable device identification in long rang (LoRa) based Internet of Things (IoT). In recent years, the identification precision of RFF has been significantly improved by leveraging artificial intelligence (AI) technologies to deeply exploit RFF features which are hardware-level, unique and resilient. However, traditional AI technologies lack strong interpretability, require massive amounts of training data and occupy huge computing resources. To address above challenges, we in this paper propose a deep fractional scattering network (DFSNet) to extract the RFF features hidden in non-stationary LoRa chirp signal through linear translation-variant multiscale fractional wavelet filters. Due to the fractional-domain deformation stability in DFSNet, the influence of noise on feature extraction can be reduced to the greatest extent by fractional transformation. Firstly, we apply DFSNet to build a hybrid RFF identification interpretability framework where the scattering coefficients of input can be calculated and characterized. Ben-efiting from the application of fractional wavelet transform, we can clearly explain the features represented by each coefficient. Then, the robustness characteristic of the fractional deformation is analyzed. Finally, experiment results show that our proposed hybrid DFSNet can achieve up to about 98.5% recognition accuracy rate with only about 5000 LoRa practical training samples per device.

Lightweight device authentication is a critical aspect in edge computing to guarantee the rightness of edge device identities and services. Radio frequency fingerprinting (RFF) is such an enabling technology able to provide robust and affordable security by employing the unique device and channel features which are usually extracted via machine learning (ML) method. The challenge is how to realize strong interpretability and support sufficient generalization ability under non-stationary channel characteristics. To solve this, we in this paper propose a novel hybrid network named FWSResNet which exploits fractional wavelet scattering transform and residual neural network to deal with the device and channel features subtlety. In particular, the proposed FWSResNet uses the scattering network based on fractional domain wavelet transform to extract the low and high-frequency features of the input signal through multiscale fractional wavelet. We find that this design can be robust to non-stationary signals and can extract the key features of noise signals. We also present a comprehensive theoretical analysis of the performance of FWSResNet under non-stationary signal distortion. Finally, we evaluate the hybrid network under largescale long term evolution (LTE) data in the practical application scenario and show that our proposed FWSResNet can achieve 93% recognition accuracy rate with only 280 training samples per device, and achieve up to 99.5% when training samples increase to 4200 per device.

With the ever-increasing growth of wireless communication technologies and the proliferation of the Internet of Things (IoT), intelligent authentication systems to distinguish legitimate devices are of vital importance. These years, deep learning based authentication algorithms have achieved considerable precision by leveraging radio frequency (RF) fingerprints. However, these methods depending on massive labeled data are difficult to apply on large-scale devices identification. In this paper, we propose a novel method using semi-supervised deep learning employing RF fingerprinting with meta pseudo time-frequency labels to improve identification performance in small-scale labeled datasets. We demonstrate how the scale of datasets and the proportion of labeled data influence the accuracy of identification by analyzing a dataset of 40 GB real Long-Term-Evolution (LTE) mobile phone’s raw signals. Experimental results show that compared with the non-convergence of traditional supervised learning with 100 labeled data, our method can achieve the authentication accuracy of 99.86% with the same labeled data. Moreover, when using the same scale of training datasets and labeled half, our approach could obtain authentication accuracy higher than traditional supervised learning. And even 1% labeled data of 900 training data, this method can still obtain the accuracy of 91.25%.

Nowadays, a large number of intelligent devices and smart sensors are being connected by various device identification and/or authentication protocols to satisfy various requirements of 5G services. However, how to identify devices by hardware-level radio frequency (RF) fingerprints of real mobile phones has been rarely researched. In this paper, we propose a novel deep learning (DL) based RF fingerprinting ResNet (RFFResNet) to identify different real mobile phones precisely by employing RF fingerprints hidden in wireless signals. Specifically, we quantitatively show how identification accuracy is influenced by channel conditions, noises, the scale of training data and network parameters. We also evaluate the proposed RFFResNet by using a dataset of 220GB long term evolution (LTE) simulation raw time data and a dataset of 25GB real mobile phone's raw time signals. Experiment results show that our RFFResNet can achieve about 95%-99% identification accuracy in real LTE application scenario and show great superiority compared with other existing DL model, such as ResNet18-1D, ResNet34-1D and VGG16-1D.