Wireless jamming identification, which detects and classifies electromagnetic jamming from non-cooperative devices, is crucial for emerging low-altitude wireless networks consisting of many drone terminals that are highly susceptible to electromagnetic jamming. However, jamming identification schemes adopting deep learning (DL) are vulnerable to attacks involving carefully crafted adversarial samples, resulting in inevitable robustness degradation. To address this issue, we propose a differential transformer framework for wireless jamming identification. Firstly, we introduce a differential transformer network in order to distinguish jamming signals, which overcomes the attention noise when compared with its traditional counterpart by performing self-attention operations in a differential manner. Secondly, we propose a randomized masking training strategy to improve network robustness, which leverages the patch partitioning mechanism inherent to transformer architectures in order to create parallel feature extraction branches. Each branch operates on a distinct, randomly masked subset of patches, which fundamentally constrains the propagation of adversarial perturbations across the network. Additionally, the ensemble effect generated by fusing predictions from these diverse branches demonstrates superior resilience against adversarial attacks. Finally, we introduce a novel consistent training framework that significantly enhances adversarial robustness through dual-branch regularization. Simulation results demonstrate that our proposed methodology is superior to existing methods in boosting robustness to adversarial samples.

Multiple-input multiple-output (MIMO) technology in conjunction with orthogonal frequency division multiplexing (OFDM) transmission is widely adopted in fifth-generation mobile networks to support multiple users. However, in these mobile communication systems, high power amplifiers (HPAs) at user terminals’ transmitters are driven into their saturation regions, which makes the multiuser frequency-selective MIMO-OFDM uplink channel nonlinear and renders the standard multiuser detection (MUD) at the base station (BS) ineffective. In this paper machine learning is employed to combat the distortions in the uplink of this multiuser frequency-selective MIMO-OFDM communication system. More specifically, a powerful complex-valued B-spline neural network (BSNN) based design is developed to simultaneously identify the system’s channel impulse response (CIR) matrix and the BSNN model for the nonlinear transmitters’ HPA together with the BSNN inversion for the nonlinear HPA at transmitters. This enables the BS to effectively implement MUD by utilizing the estimated MIMO-OFDM CIR matrix as well as to compensate for the transmitter HPAs’ saturation distortions using the estimated BSNN inversion. A simulation study is included to evaluate the effectiveness of this novel BSNN assisted design in combating multiuser and dispersive channel interference as well as nonlinear distortions for multiuser MIMO-OFDM nonlinear uplink.

Deep learning based channel prediction has garnered significant attention to mitigate channel aging in high-mobility multiple-input multiple-output (MIMO) systems. However, existing channel prediction methods extract the temporal correlations from the channel sequences estimated at uniform pilots, which require dense pilot configuration to mitigate Doppler aliasing in high-mobility scenarios and incur substantial estimation overhead. To tackle this problem, we propose a channel prediction method based on continuous-time transformer with the non-uniform pilot pattern, thereby enabling accurate prediction across arbitrary time scales with only a small number of pilots. Specifically, we first design the non-uniform pilot pattern based on Chebyshev polynomial roots and then prove its optimality under Doppler-dominated channel variations with relatively stable user velocity, wherein a subset of pilots are densely configured to provide a finer resolution of Doppler phase estimation. To adapt to the non-uniform pattern, a continuous-time transformer is further proposed, which integrates the superior feature extraction capability of transformer with the continuous-time modeling strength of neural ordinary differential equation (ODE) for flexibly processing the estimated channel sequences with non-uniform time scales. More concretely, the attention mechanism is extended to the continuous-time domain by incorporating neural ODE, while a high-frequency temporal encoding is designed to fit rapidly time-varying channels. Besides, an element-wise prediction mechanism is proposed to efficiently capture temporal correlations and prevent overfitting. Simulation results demonstrate that our proposed method can realize accurate continuous-time channel prediction in high-mobility scenarios, and significantly outperforms existing channel prediction methods.

In unlicensed spectrum, wireless communication systems are vulnerable to electromagnetic attacks and interference from non-cooperating parties, thus amplifying the significance of wireless communication security. The identification of wireless interference serves as a pivotal technology for spectrum sensing and is crucial to facilitate anti-interference communications, where wireless interference identification adopting deep learning technology has been extensively explored and has exhibited exceptional performance benefits. In this paper, we propose a lightweight transformer network (LTN) for interference identification, which can solve the computational complexity challenge from the conventional transformer networks while preserving their global feature extraction proficiency. LTN comprises three lightweight modules, namely low-complexity linear embedding (LCLE), integral and refined feature extraction (IRFE) and attention matrix reuse (AMR). Firstly, the LCLE module is obtained through the utilization of reparameterization techniques. The incorporation of reparameterization enables the decoupling of the network architecture during the training and testing phases, thereby enhancing the performance and reducing the complexity concurrently. Secondly, we propose the IRFE module, which leverages the discrete wavelet transform to partition the input into integral and refined components. For feature extraction, multi-head self-attention (MSA) is utilized for the refined part while window-based MSA is employed for the integral part, ensuring an optimized allocation of computational resources. Finally, we present a novel AMR mechanism, which takes advantage of the similarity of attention matrices in adjacent MSA layers. AMR can effectively circumvent the computational complexity by saving subsequent attention matrix computations. Simulation results validate that our proposed methodology has higher recognition accuracy.

Benefiting from a large amount of unallocated bandwidth, millimeter-wave (mmWave) communications have been regarded as one of the most promising technologies. To overcome high pathloss of mmWave signals, the beamforming technique plays a fundamental role. In recent years, with the success of deep learning (DL), DL-based beam prediction methods have been widely studied to reduce the training overhead of traditional beam scanning methods. In this paper, a novel DL-based low-overhead beam prediction scheme is proposed, which is motivated by two important observations: (1) The optimal beam prediction is difficult for non-line of sight (NLOS) scenario, which limits the overall prediction accuracy. (2) On the contrary, the optimal beam can be precisely predicted with low computational costs under line of sight (LOS) scenario. Therefore, we propose a flexible network architecture, namely multi-stage network (MSN), to conduct the optimal beam prediction. Firstly, MSN contains multiple branches with gradually increasing computational complexity, and each branch carries with a classifier, which enables the MSN to have the capability of adaptively and dynamically allocating computational resources. Meanwhile, to combine the advantages of convolutional neural network (CNN) and transformer for feature extraction in MSN, we design joint CNN and transformer (JCT) module and its simplified module, namely Ghost-JCT. Secondly, we propose two pre-training strategies to effectively improve the performance of classifiers without additional computational costs. Finally, we propose confidence-based and Markov-based classifier selection strategies, which could select the appropriate classifier to strike a balance between accuracy and computational complexity. Simulation results demonstrate that MSN enjoys significant superiority in terms of computational complexity and prediction accuracy compared to its traditional counterparts.

In non-cooperative communications, malicious electromagnetic interference attacks communication systems and causes higher probability of communication disruption. In order to address the challenges posed by electromagnetic interference, the wireless interference recognition technique has emerged, which identifies the interference signals without priori information. In recent years, the success of deep learning (DL) has sparked interest in introducing DL in the field of wireless interference recognition. However, most DL-based interference identification methods improve accuracy by dramatically increasing network sizes while ignoring the important effect of network inputs. For this reason, we extensively investigate the impact of different signal transformation forms of interference (called signal modalities) on performance. The artificial features of the interference signal are also utilized as one of the refined modalities, which breaks the inherent concept that artificial features are only used in the methods of feature extraction. Convolution and transformer are combined in the extraction of different modal features. In order to reduce the complexity of transformer, a dual transformer module (DTM) is proposed. Furthermore, to overcome the imbalance of modal optimization during the training process, an adaptive gradient modulation (AGM) strategy is proposed, which leads to better convergence for the multimodal training. Finally, modal information selection mechanism (MISM) selects the most appropriate modalities for each input sample, which saves computational costs. Extensive experiments demonstrate that combining multiple interference modalities is more effective than trying different networks.

As electromagnetic environments are more and more complex, wireless interference identification (WII) is becoming vital for non-cooperative communication systems in both civilian and military scenarios. With the enormous success of deep learning (DL), methods that optimize convolutional neural networks (CNNs) for WII have been proposed. However, due to the intrinsic characteristics of CNNs, the existing networks are difficult to capture long-range feature dependencies, causing the low recognition accuracy and the high computational complexity. Motivated by the success of transformers in natural language processing (NLP) domain, we propose an augmented convolutional neural network with transformer (ACNNT), which combines both the advantages of CNNs and transformers to simultaneously strengthen locality and establish long-range dependencies. Specifically, the ACNNT has multiple stages, and every stage consists of convolutional layers and transformer module to model local and long-range dependencies of context, respectively. At the end of the network, a classification token is used for classification. A channel attention (CA) module is proposed to further improve the expressive ability of the transformers. Extensive experiments demonstrate that the proposed method leads to performance improvement as compared to conventional DL based methods.

Wireless interference identification (WII) is a promising technology for non-cooperative communication systems in both civilian and military scenarios. With the rapid development of artificial intelligence, deep learning (DL) based WII methods have been proposed. However, the existing networks based on DL do not have the adaptive ability, and this situation may cause a waste of computing resources. In this paper, we propose a novel Multi-Depth Adaptive Network (MDANet), and it can adaptively determine the forward propagation depth according to the difficulty of input samples in inference. Specifically, we firstly divide a given convolutional neural network (CNN) into several blocks, and each block has its own output of classification by attaching a fully connected layer. The mechanism of confidence is introduced to enable the network to dynamically select depth of forward propagation and allocate appropriate computational resources depending on the complexity of samples during test time. In addition, we improve the recognition ability of early blocks through three proposed approaches. Experiments demonstrate that the MDANet can reduce the calculation cost and prediction time significantly without the performance loss.