Luxin Zhang

The adaptability of modulation recognition in open environments is challenging, making unsupervised modulation clustering an increasingly prominent research focus. However, traditional unsupervised clustering methods that rely only on data-driven learning have difficulty maintaining stable clustering performance under complex channel conditions. In this paper, we propose a knowledge-enhanced contrastive clustering (KECC) method for unsupervised signal modulation clustering, which introduces a new paradigm of clustering guided by expert knowledge in signal processing. KECC consists of two stages: first, in the knowledge-enhanced contrastive pre-training stage, expert knowledge is embedded into the shared feature extraction network by constructing an auxiliary expert knowledge branch and a data augmentation branch. Then, in the knowledge-screening contrastive fine-tuning stage, the neighborhood relationship guided by expert knowledge is utilized for feature optimization, and a clustering head is directly trained to achieve unsupervised clustering. Experimental results on multiple public modulation datasets show that KECC significantly outperforms existing baseline methods in terms of ACC, ARI and NMI metrics, achieving state-of-the-art clustering performance.

Deep learning-based wideband spectrum sensing models have achieved impressive performance in subband occupancy detection, but their susceptibility to adversarial attacks raises serious concerns. In this paper, we construct a targeted adversarial attack framework specifically designed for multi-label wideband spectrum sensing tasks. We further propose two optimization-based attack methods. MLS CW formulates perturbation generation as a constrained optimization problem and solves it via gradient descent. MLS DF is a greedy iterative method that approximates the solution to an underdetermined linear system through linearization. Both attacks are developed under strict power constraints and are capable of generating subtle yet highly effective perturbations. These attacks further incorporate a threshold-aware loss and selective subband targeting to enhance precision and stealthiness. Extensive experiments across various SIR conditions demonstrate that the proposed attacks significantly degrade the model’s detection performance in both untargeted and targeted scenarios, while maintaining high stealthiness. These results expose the severe vulnerability of DL-based spectrum sensing models and highlight the need for further research into their robustness under adversarial settings.

Despite significant advancements in deep learning-based DOA estimation, such models exhibit critical vulnerabilities to adversarial attacks. Previous studies of attacks on DOA estimation models typically require full knowledge of the target model, which is impractical in real-world black-box scenarios. In this paper, we explore two black-box adversarial attack strategies for deep learning-based DOA models: transfer-based and decision-based attacks. Decision-based attacks iteratively generate adversarial samples by querying the model’s hard-label predictions, while transfer-based attacks craft perturbations using surrogate models with accessible architectures and data. We systematically investigate how the architecture, depth, and input of the surrogate model affect attack transferability. Under an 8-element uniform linear array setting, experimental results reveal that both strategies can substantially degrade the performance of deep learning-based DOA models and reach attack success rates above 88% in black-box settings. Crucially, we further demonstrate that adversarial samples designed for deep learning models also impair classical subspace-based DOA estimation methods, yielding RMSE values that are 1.39 to 50.77 times higher than those under AWGN, depending on the attack type.

This paper proposes MF-Pnet, a deep learning-based positioning method designed to address the limited utilization of informative features in existing FM signal positioning approaches. Unlike prior methods that process the wideband FM signal as a whole, MF-Pnet extracts and fuses multiple narrowband IQ signals that carry potentially complementary information across different frequency bands. A multi-branch parallel feature extraction network is constructed to independently model the features from each narrowband signal, and an attention mechanism is introduced to adaptively weight their relative contributions. This design significantly enhances the overall feature representation and improves positioning accuracy. Comprehensive experiments conducted on a real-world FM dataset demonstrate that MF-Pnet outperforms the state-of-the-art FM-Pnet in both indoor and outdoor environments, exhibiting significant advantages in terms of accuracy, stability, and cross-day generalization capability. Furthermore, MF-Pnet maintains practical inference efficiency, making it well suited for deployment on resource-constrained devices.

Deep learning has achieved remarkable success in direction-of-arrival (DOA) estimation. However, recent studies have shown that adversarial perturbations can severely compromise the performance of such models. To address this vulnerability, we propose Transformer-based Adversarial Defense for DOA estimation (T-ADD), a transformer-based defense method designed to counter the evaluated adversarial attacks. To achieve a balance between robustness and estimation accuracy, we formulate the adversarial defense as a joint reconstruction task and introduce a tailored joint loss function. Experimental results demonstrate that T-ADD mitigates the evaluated PGD and MIM attacks in far-field narrow-band scenarios. Compared to existing defense methods, the proposed approach achieves RMSE reductions of up to 83% under the evaluated experimental settings.

To overcome the limitations of traditional satellite navigation technologies in complex and signal-obstructed industrial environments, this article presents MSM-Pnet, a novel semi-supervised frequency modulation (FM)-based positioning framework leveraging FM signals of opportunity (SOPs). By integrating wavelet packet decomposition (WPD) with a multiscale Vision Transformer (MS-ViT) and a hybrid masking strategy that combines random and time–frequency-aware masking, MSM-Pnet introduces a masked autoencoder (MAE) architecture capable of robust positioning with limited labeled data. Experimental results demonstrate that MSM-Pnet consistently outperforms conventional supervised learning methods in both indoor and outdoor environments, while also significantly reducing model complexity. These results highlight the method’s potential as a cost-effective and scalable solution for seamless indoor–outdoor positioning for Internet of Things (IoT) systems.