Phishing poses a significant threat to the financial and privacy security of internet users and often serves as the starting point for cyberattacks. Many machine-learning-based methods for detecting phishing websites rely on URL analysis, offering simplicity and efficiency. However, these approaches are not always effective due to the following reasons: (1) highly concealed phishing websites may employ tactics such as masquerading URL addresses to deceive machine learning models, and (2) phishing attackers frequently change their phishing website URLs to evade detection. In this study, we propose a robust, multi-view Transformer model with an expert-mixture mechanism for accurate phishing website detection utilizing website URLs, attributes, content, and behavioral information. Specifically, we first adapted a pretrained language model for URL representation learning by applying adversarial post-training learning in order to extract semantic information from URLs. Next, we captured the attribute, content, and behavioral features of the websites and encoded them as vectors, which, alongside the URL embeddings, constitute the website’s multi-view information. Subsequently, we introduced a mixture-of-experts mechanism into the Transformer network to learn knowledge from different views and adaptively fuse information from various views. The proposed method outperforms state-of-the-art approaches in evaluations of real phishing websites, demonstrating greater performance with less label dependency. Furthermore, we show the superior robustness and enhanced adaptability of the proposed method to unseen samples and data drift in more challenging experimental settings.

Convolutional Neural Networks (CNNs) and Transformer are two powerful representation learning techniques for visual tracking. Although CNNs can effectively reduce local redundancy via small-neighborhood convolution operations, their limited receptive fields make it difficult to capture global dependency. Self-attention in Transformer uses patches as the input representation, which can effectively capture long-range dependency. However, blind similarity comparisons between all patches can lead to high redundancy. Is there then a technique that combines well the advantages of both paradigms for visual tracking? In this work, we design a novel backbone network for feature extraction. First, we choose Depthwise Convolution and Pointwise Convolution to build a Convolution Mixer, which effectively separates spatial mixing from channel-wise mixing of information. The Convolution Mixer reduces redundancy in spatial and channel features while increasing receptive field. Then, to exploit the global modeling ability of self-attention, we construct a module by aggregating Convolution Mixer and self-attention. The module shares dominant computational complexity (the square of the channel size) in the first stage. In the second stage, the shift and summation operations are lightweight. Finally, to alleviate the overfitting of the backbone network during training, a dropout layer is added at the end of the module to improve the generalization ability of the network model. Stronger image features are provided for subsequent feature fusion and prediction. The proposed tracker (named CMAT) achieves satisfying tracking performance on ten challenging datasets. In particular, CMAT achieves a 64.1% AUC on LaSOT and a 68.9% AUC on UAV123 while running at 23 frames per second (FPS).

Clutter removal in ground-penetrating radar (GPR) based on deep learning has been studied in recent years. However, existing methods are primarily designed for homogeneous background conditions and utilize only local spatial information via the convolution operation. In order to solve these issues, a subspace projection attention network is proposed for GPR heterogeneous clutter removal in this paper. Firstly, a heterogeneous concrete dataset based on a numerical model with randomly placed aggregates is constructed, which incorporates the complex electromagnetic propagation process accurately to improve the effectiveness for heterogeneous clutter removal. In addition, the Clutter Basis learning neural Network (CBNet) is designed by integrating the subspace projection attention (SPA) module into the skip connection paths of U-Net architecture. By learning the subspace basis vectors adaptively, the SPA exploits both local and global spatial information to extract target features precisely. At the same time, the feature maps are projected to the target subspace to remove heterogeneous clutter features. Finally, the performance and effectiveness of proposed method is validated by simulations and experiments.

The common middle point (CMP) ground penetrating radar (GPR) utilizes the time-delay information under different antenna separations to realize layered inversion. However, parameter inversion of multilayered media is mostly subject to heavy computational complexity and worst estimation error in the determination of refraction positions. To address the problems, an effective parameter inversion method is proposed based on off-grid sparse CMP model with refined orthogonal matching pursuit (OMP) for multilayered parameter inversion. In the proposed method, a sparse CMP signal model with accurate reflected wave propagation modeling is developed based on a refraction approximation method. An off-grid sparse CMP model is further constructed based on second-order Taylor expansion to overcome the deviation from the grid node. Then, a refined OMP algorithm based on compressed sensing (CS) is proposed with an off-grid optimization process to achieve accurate off-grid parameter inversion. Finally, the effectiveness of the proposed method is verified by simulations and experiments.

Modulation recognition of communication signals is of great importance in the context of the Internet of Everything (IoE), as wireless communication technology is a key foundation for implementing the IoE. Recently, graph neural networks (GNNs) have been successfully applied to modulation recognition tasks due to their ability to merge messages transmitted between adjacent nodes in the graph. However, GNN-based models are more computationally intensive when processing long signals, potentially reducing their practicality. In this article, we explore a novel signal representation from a graph perspective and propose a graph-powered modulation recognition framework. We first propose the dynamic time series graph (DTSG) algorithm, which segments the signals and maps each segment into a patch graph, with corresponding patches from different signals sharing connected edges. By integrating DTSG with both GNNs and recurrent neural networks (RNNs), we have designed an end-to-end signal classification framework, DTSG-Net, for modulation recognition. Experimental results on four datasets: 1) RML2016.10a; 2) RML2018.01a; 3) Sig2019-12; and 4) HKDD_AMC36—demonstrate that our DTSG-Net can achieve high signal modulation classification accuracy (Acc) with minimal computational resources, outperforming existing methods based on signal graph representation in terms of computational resource savings and higher accuracy.

Community detection has been a subject of extensive research due to its broad applications across social media, computer science, biology, and complex systems. Modularity stands out as a predominant metric guiding community detection, with numerous algorithms aimed at maximizing modularity. However, modularity encounters a resolution limit problem when identifying small community structures. To tackle this challenge, this paper presents a novel approach by defining community structure information from the perspective of encoding edge information. This pioneering definition lays the foundation for the proposed fast community detection algorithm CSIM, boasting an average time complexity of only $O\left(n\text{log}n\right)$. Experimental results showcase that communities identified via the CSIM algorithm across various graph data types closely resemble ground truth community structures compared to those revealed via modularity-based algorithms. Furthermore, CSIM not only boasts lower time complexity than greedy algorithms optimizing community structure information but also achieves superior optimization results. Notably, in cyclic network graphs, CSIM surpasses modularity-based algorithms in effectively addressing the resolution limit problem.