Graph Neural Networks (GNNs) have demonstrated superior performance in processing centralized graph-structured data. However, real-world privacy and security concerns hinder data centralization and shareing, leading to severe data isolation (data silos). While Federated Learning (FL) offers a distributed solution to mitigate these obstacles, existing Federated Graph Neural Network (FedGNN) frameworks struggle to effectively address data heterogeneity. To address this, this paper proposes DA-DFGAS, a federated graph neural architecture search algorithm. Specifically, DA-DFGAS facilitates model personalization via a directed tree topology and path constraint mechanisms, while simultaneously employing a joint self-attention mechanism based on predicted probability distributions to capture distributional variations across multiple clients. Furthermore, it integrates a bi-level global-local objective optimization strategy to ensure global model consistency while preserving local adaptability. Experimental results on multiple datasets demonstrate that DA-DFGAS outperforms state-of-the-art methods, achieving 0.5–3.0% accuracy improvements over centralized baselines and 0.5–5.0% over federated counterparts.

Graph Anomaly Detection (GAD) focuses on identifying instances that deviate from normal patterns in graph-structured data. Although substantial progress has been made in this field, current approaches are constrained by the "one-dataset-one-model" paradigm, exhibiting limited generalization across heterogeneous graphs, poor adaptability in few-shot scenarios, and inefficient cross-domain deployment. To overcome these limitations, we propose SAARCS, a universal GAD framework capable of performing anomaly detection across diverse graph datasets without requiring any target data training. SAARCS aligns feature dimensions through composite spatial smoothness, learns graph embeddings via an adaptive-hop attention encoder, and predicts node abnormality using only a small set of normal context nodes. Extensive experiments on eight real-world datasets demonstrate that our approach achieves superior performance compared to state-of-the-art baselines.

Multimodal Sentiment Analysis (MSA) and Emotion Recognition in Conversations (ERC) have rapidly developed into pivotal tasks in artificial intelligence. Large Language Models (LLMs) offer powerful semantic reasoning and computational capabilities, showing great potential for understanding emotional content. However, when applied to multimodal sentiment data, LLMs face significant challenges, including the inability to directly process heterogeneous data, difficulties in coping with feature misalignment and suboptimal cross-modal fusion. To address these challenges, we propose a novel multimodal sentiment inference framework named ATGFB-MFF which grounded in fiber bundle theory. This method decomposes multimodal features into an adaptive text-guided shared semantic space and fiber offset spaces to achieve structured alignment and fusion. Then the fused features are converted into structured pseudo-token sequences for effective inference via frozen LLMs. We also introduce two loss functions respectively called shared space consistency loss and fiber offset regularization loss which are used to improve representation stability. Extensive experiments on four benchmark datasets demonstrate that ATGFB-MFF consistently outperforms state-of-the-art baselines. These results highlight the efficacy of geometric structural modeling in unlocking the potential of LLMs for multimodal sentiment inference.

Data-driven spectrum sensing is a key technology for addressing complex challenges in Cognitive Radio Networks (CRNs). Traditional methods are typically designed for simple single-band scenarios and perform poorly in practical wideband applications. In real-world systems, a single Secondary User (SU) is often restricted by energy, time, and hardware capabilities during real-time sensing. Consequently, only local and fragmented frequency information can be obtained. This partial sensing leads to a severe lack of training data. Additionally, the lack of historical records for emerging frequency bands, combined with data incompleteness due to resource constraints, creates training bottlenecks for data-driven models and limits the reliability of sensing. To address these challenges, this paper proposes a novel framework based on Collaborative Subgraph Learning and Hyperbolic Graph Neural Networks (GNNs). This approach enables Secondary Users to perform collaborative sensing through distributed subgraph learning. By utilizing GNNs to extract features and model multi-band correlations, a new distributed GNNs architecture is designed to efficiently detect wideband spectrum occupancy, even with partial observations. Within this framework, all frequency bands in the wideband spectrum pool are treated as a unified graph, while the bands observed by each SU form a subgraph. Subsequently, the complete spectrum graph is constructed through the joint training and aggregation of these subgraphs. By integrating hyperbolic geometry into GNNs, this method better captures the hierarchical structure of spectrum patterns, providing a more accurate and efficient sensing model. Experimental results demonstrate that, compared to the second-best HCNNs model, the proposed framework improves sensing accuracy by 3.8% on average across various test environments, while reducing key resource consumption by 18.4% on average.

In recent years, cross-domain speech emotion recognition (SER) has attracted considerable interest. Most transfer subspace learning based SER methods lack unified adaptive constraints, making it difficult to balance discriminative capability and domain alignment, which limits their cross-domain generalization. To address these problems, we propose a novel domain adaptation (DA) approach called dynamic graph consistent weighted subspace learning (DGCWSL). Specifically, DGCWSL first projects samples from the source and target domains into a shared low-dimensional discriminative subspace, then performs cross-domain instance reconstruction, representing each target as a weighted combination of source instances. In parallel, a dynamic graph is constructed to capture local structural information between domains while preserving the data manifold. Subsequently, label supervision and discriminative learning between domains are achieved through linear regression. Furthermore, we introduce an adaptive weighted matrix that enforces consistent feature contributions across the distance metric, instance alignment, and discriminative regression, thereby mitigating both overfitting and underfitting. Finally, extensive experiments are conducted on four public datasets. The results confirm the superiority of DGCWSL over several state-of-the-art DA methods.

Spectral–spatial feature modeling plays a crucial role in hyperspectral image (HSI) classification. However, existing models based on convolutional neural networks (CNNs) and Transformers still face a trade-off between feature modeling capability and computational efficiency. Although recent wavelet-based HSI classification methods have demonstrated the advantages of frequency-domain analysis, they typically rely on a single wavelet basis, which limits their ability to capture diverse spectral–spatial patterns across different frequency bands. To address these issues, we propose Wavelet Spectral-Spatial Mamba (WSSMamba) by combining wavelet transform with state space modeling for HSI classification. WSSMamba introduces an Adaptive Wavelet Fusion Module (AWFM) to perform multi-scale frequency domain decomposition using multiple wavelet bases. This allows the model to extract both low-frequency global structure and high-frequency local details. A Wavelet Feature Enhancement (WFE) module is also designed to improve feature discriminability by applying channel and spatial attention mechanisms. Furthermore, we propose a Spectral-Spatial Cross-Fusion Strategy (SSCFS), which uses multi-directional state modeling to dynamically integrate high-frequency information. Extensive experiments on benchmark datasets demonstrate that WSSMamba outperforms state-of-the-art methods in classification performance.

Currently, multimodal sentiment analysis (MSA) for personalized users under uncertain modalities missing has become a new challenging problem. To address this issue, we propose a two-step idea. First, we propose an effective MSA model under uncertain modalities missing and train it with some public datasets, thus to enable the model to possess better preliminary MSA ability. Then, we make the pretrained model to continuously learn user’s personalized characteristics with online learning methods, thereby enable the model grow into a robust model for personalized MSA. Based on this idea, we propose a Personalized MSA model under uncertain modalities missing via Pretraining and Online Learning (termed as PMSAPO). For Personalized MSA under uncertain modalities missing, PMSAPO firstly generates the fused modality and allocate weights for each modality with a Fully Connected Neural Network Evaluation Module. Then, PMSAPO completes the final sentiment classification based on the fusion modality with a Joint feature optimization module. For the pretrained PMSAPO, we make it autonomously learn the personalized users via our proposed online learning techniques, including an online meta-learning method, a learning rate adaptive adjustment strategy, and a dynamic weight assignment strategy for sample data. Finally, based on three public benchmark datasets (IEMOCAP, MELD and CMU-MOSI), we conduct extensive experiments and prove that PMSAPO completely outperforms the Twelve state-of-the-art baseline models. (Code is available at https://github.com/SHX-AI/PMSAPO .)

With the popularity of smart mobile devices embedded with rich sensors, mobile crowdsensing (MCS) has gradually attracted the attention of researchers in recent years. Task allocation is a key research problem in MCS systems, where platforms recruit workers and assign them crowd tasks. While previous research has focused on the utility of recruiting workers, the location factor of workers has been ignored. Therefore, this paper proposes a two-stage worker recruitment framework named BW-Selector, which recruits workers in two stages. In the offline stage, this paper proposes an opportunity-crowd worker recruitment algorithm, which first divides the task area with a Voronoi diagram, and then builds a prediction model based on long short-term memory (LSTM) to predict the movement trajectory of workers and solve the cold start in the traditional MCS system. In the online stage, for maximizing the task space coverage under the premise of a limited task budget, this paper proposes a participatory-crowd worker recruitment algorithm based on adaptive threshold selection. Finally, through experiments on real datasets, it is verified that BW-Selector has better performance in terms of task space coverage and running time under the same constraints compared with other methods.

Cross-domain speech emotion recognition (SER) has received increasing attention in recent years. Existing transfer subspace learning and regression-based SER methods have the following drawbacks. The features in the subspace are still insufficiently representative and discriminative, and direct regression would lead to information loss. To address these problems, we present a novel common discriminative latent space learning (CDLSL) method for cross-domain SER. To be specific, we first obtain a common latent space by imposing a projection matrix on the cross-domain data. Meanwhile, we impose an uncorrelated constraint on the projection matrix to ensure that the features are representative and discriminative after dimension reduction. Then, we implement a graph regularization term on the latent representations of the samples to capture the local similarity information. Furthermore, to obtain a more discriminative common latent space, we introduce the label information by aligning the latent space with the relaxed label space, while mitigating the information loss for regression. Extensive experimental results validate the superiority of the proposed method over the state-of-the-art competitors.

The industrial internet of things (IIoT) plays a crucial role in manufacturing, logistics, and equipment management. Graph neural networks (GNNs) can effectively model graph-structured data and have received widespread attention in IIoT applications. However, existing methods face key challenges. First, IIoT data typically contains sensitive information, making it difficult to conduct centralized training on dispersed data. Second, the current model fails to fully capture the complex interrelationships between different devices. To address the above issues, this article proposes a federated learning-based graph neural network model FedMRGNN for joint analysis of distributed IIoT. This model performs federated learning through model aggregation and parameter exchange, while protecting privacy through differential privacy mechanisms. Meanwhile, to better capture the complex relationships between devices, this article integrates multiscale feature extraction and residual connections into the model. Multiscale feature extraction can process graph data in parallel through multiple branches, each branch using convolutional kernels of different scales to extract node features, and utilizing multiscale pooling operations for local aggregation and dimensionality reduction. Residual connections can enhance the fusion ability of multiscale features and alleviate the problem of gradient vanishing in deep network training. In order to further verify the effectiveness of FedMRGNN, experimental verification was conducted on different datasets. The results show that FedMRGNN improves classification accuracy by 2.21%–6.40% compared to other baseline algorithms in most scenarios. In practical IIoT applications, improved classification accuracy can help predictive maintenance systems detect potential device failures in advance, thereby improving overall device operational efficiency.

With the rapid development of the Ethereum platform, phishing fraud has become increasingly rampant, posing significant security risks to both users and the platform. However, existing phishing fraud detection methods are manually designed, requiring substantial human effort, and are unable to adapt to diverse detection scenarios. In this article, we propose phishing detection on Ethereum via graph neural architecture search of transaction subgraph (PETS-GNAS). The phishing detection problem on Ethereum is transformed into a graph classification task, where accounts and transactions are represented as nodes and edges, respectively. Specifically, we acquire account labels and their corresponding transaction information from credible sources and then extract transaction subgraphs centered on labeled accounts as datasets. Subsequently, we introduce a mapping mechanism to extend these transaction subgraphs into corresponding temporal transaction subgraph (TTSG), encoding transaction attributes during the TTSG construction process. Then, graph neural architecture search (GNAS) strategy that incorporates early stopping and L2 regularization is proposed to enhance the feasibility and accuracy of Ethereum phishing detection by avoiding redundant parameters and complex architectures. Extensive experimental results demonstrate that PETS-GNAS achieves strong performance in phishing detection tasks, enabling early and accurate identification of phishing accounts.

High resolution optical remote sensing images are the guarantee for remote sensing image analysis and application. However, many images suffer from blurring, distortion and low resolution due to camera hardware limitations and unstable image transmission. To address these challenges, we propose a Random Degradation and Adaptive Fusion-based super-resolution network (RDAF-GAN) for improving the clarity and detail of images. Specifically, unlike the traditional single degradation method, we design a comprehensive simulation model for remote sensing image degradation. It aims to generate low-resolution remote sensing images that are closer to the real scene. Subsequently, these generated low-resolution images are fed into RDAF-GAN for reconstruction to recover finer and more accurate image details. In addition, we propose an image fusion method based on local contrast. By adaptively adjusting the fusion weights, the perceived clarity and visual quality of the images are further enhanced. The experimental results validate that RDAF-GAN outperforms other state-of-the-art methods and consistently produces excellent results in a variety of situations.

Lane detection is crucial for autonomous driving systems (ADS), utilizing sensors like cameras and LiDAR to identify lanes and understand vehicle position, direction, and lane shape. It provides data support for the control system to make informed driving decisions. In this survey, we review recent advancements in lane detection, focusing on both 2D techniques and emerging 3D methods. We begin with an overview of the significance of lane detection in ADS, followed by an analysis of the evolution of 2D techniques over the past decade, covering traditional and deep learning approaches. We also examine recent advancements in 3D lane detection. Additionally, we summarize evaluation metrics and popular datasets in the field. Finally, we discuss current challenges and future directions in lane detection, aiming to provide valuable insights for researchers and developers in this technology.

3D lane detection provides richer spatial information than 2D lane detection planar position results. It improves vehicle perception in complex scenes, which is becoming increasingly important in intelligent driving. However, existing frameworks mainly focus on mapping front-view (FV) and bird’s-eye view (BEV) features and ignore the intrinsic correlations between different perspectives and scales. It can lead to incomplete feature extraction, affecting the perception accuracy of lane detection and adaptation ability to complex scenes. To alleviate these problems, we present a novel Freq-3DLane framework, an efficient end-to-end 3D lane detector. Instead of directly superimposing deeper and lower-level features, we propose a strategy for multi-scale information integration that exploits the frequency characteristics of features for image feature extraction. To enhance perception, we fuse image features at each scale through frequency processing to ensure that detailed information and global structure are fully utilized. Next, spatial transformation fusion captures the association between the FV and the BEV feature at any two-pixel position of both, thus enabling view feature transformation. In addition, attentional guidance enhances the lane semantic information to ensure recovery of the lane geometry for accurate 3D lane detection. Extensive results on two challenging benchmarks (Apollo 3D Lane Synthetic, and OpenLane) show that our model performs favorably against the state of the arts.

With the rapid progression of mobile crowdsourcing (MCS) technology, its growing influence in our daily lives has established it as a crucial component of modern society. However, while the convenience of MCS is widely appreciated, it also poses significant threats to personal privacy, particularly location privacy. This article introduces a novel system for personalized location privacy protection in MCS. The system is divided into three main parts. The first part presents an innovative algorithm that calculates the location privacy level of crowd workers. This algorithm is crucial in determining the location privacy level required by each individual crowd worker. The second part involves the design of a personalized differential privacy protection (P-DP) algorithm, which is based on the exponential mechanism. This algorithm provides varying degrees of privacy protection strength, tailored to the location privacy protection level of each crowd worker. Furthermore, we incorporate a trusted third party (TP) server to act as an intermediary. This server eliminates any correlation between the crowd workers and the data. It is also tasked with calculating the location privacy level and reward for each crowd worker. The third part of the system is the personalized localized differential privacy (LDP) protection (P- LDP) algorithm, this algorithm is designed to further solve the problem of privacy disclosure caused by the TP server being attacked. Finally, we have conducted a comprehensive evaluation of the proposed location privacy protection system using real data sets, and the results demonstrate that the system can effectively balance the location privacy protection of crowd workers and the availability of location data, thereby improving the efficiency and reliability of MCS.

With the development of the Industrial Internet of Things (IoT), the work of large-scale data collection makes spatiotemporal crowdsensing (SC) play an important role. Mobile devices equipped with sensors could act as workers to collect and process data for uploading. In the task allocation process, a fully static allocation fails to meet the needs of realistic conditions, while a completely dynamic allocation fails to achieve the desired results. Therefore, we assume a task-scheduled execution scenario that combines the above two conditions. In the pre-allocation process, an original time location constraints (ORTA) allocation algorithm is first proposed. Then it is optimized (OPTA) to fully utilize the remaining time of the workers and increase the matched number. In addition, the design of the incentive mechanism is an effective means to improve the task completion rate of the platform. To efficiently utilize the limited platform budget in the long run, a Q-learning-based algorithm is proposed to identify target inspire tasks and subsequently increase their reward to attract workers’ participation. Finally, comparison experiments are conducted on real datasets to verify the effectiveness of our algorithm. Furthermore, the experiments on a Raspberry Pi local terminal are conducted under a satellite-based environment.

Mobile crowd sensing (MCS) is an emerging approach to collect data using smart devices. In MCS, task assignment is described as assigning existing tasks to known workers outside the constraints of task demand attributes and worker attributes, and maximizing the profit of the platform. However, workers and tasks often exist in different environments and heterogeneous features such as workers with attributes are not considered, leading to nondeterministic polynomial (NP)-hard task assignment problems. To optimize such problems, this article proposes a multiattribute environments-classes, agents, roles, groups, and objects (E-CARGO) task assignment model based on adaptive heterogeneous residual networks (AHRNets). The AHRNet is integrated into deep reinforcement learning (DRL) to optimize the NP-hard problem, dynamically adjust task assignment decisions and learn the relationship between workers with different attributes and task requirements. Multiattribute E-CARGO uses group task assignment policy to obtain the ideal worker-task assignment relationship. Compared with traditional heuristic algorithms for solving NP-hard, this method has the flexibility and applicability of adaptive networks, enabling the solver to interact with and adapt to new environments and generalize its experience to different situations. Under various experimental conditions, a large number of numerical results show that this method can achieve better results than the reference scheme.

Mobile crowdsourcing (MCS) is a new paradigm that uses various mobile devices to collect sensed data. Mobile edge computing (MEC) can effectively utilize the device resources of mobile edge, greatly relieve the pressure of network bandwidth and improve the response speed. In this article, we construct a four-party evolutionary game model consisting of the platform, crowd workers, task requesters, and edge servers. The computing tasks are conducted on edge servers, which greatly reduce remote data transmission and network operating costs and improve service quality. Taking into account the collusion between the platform and workers, and that between the platform and requesters, we analyze the stability of the strategic equilibrium in MCS using replicator dynamics methods. The optimal payoff strategies of the participants in different initial states are obtained. To prevent cheating and false-reporting problems, reward and punishment strategies are provided. Finally, the stability of the equilibrium of the four-party evolutionary game system is verified by simulation experiments, and an incentive strategy is designed to motivate all parties to choose the trust strategies.

Collaborative Mobile Crowdsourcing (CMCS) allows platforms to recruit worker teams to collaboratively execute complex sensing tasks. The efficiency of such collaborations could be influenced by trust relationships among workers. To obtain the asymmetric trust values among all workers in the social network, the Trust Reinforcement Evaluation Framework (TREF) based on Graph Convolutional Neural Networks (GCNs) is proposed in this paper. The task completion effect is comprehensively calculated by considering the workers' ability benefits, distance benefits, and trust benefits in this paper. The worker recruitment problem is modeled as an Undirected Complete Recruitment Graph (UCRG), for which a specific Tabu Search Recruitment (TSR) algorithm solution is proposed. An optimal execution team is recruited for each task by the TSR algorithm, and the collaboration team for the task is obtained under the constraint of privacy loss. To enhance the efficiency of the recruitment algorithm on a large scale and scope, the Mini-Batch K-Means clustering algorithm and edge computing technology are introduced, enabling distributed worker recruitment. Lastly, extensive experiments conducted on five real datasets validate that the recruitment algorithm proposed in this paper outperforms other baselines. Additionally, TREF proposed herein surpasses the performance of state-of-the-art trust evaluation methods in the literature.

Video stabilization can improve the visual quality of videos that have been captured on mobile devices or other handheld cameras, which are more prone to shaking and motion artifacts. Most of the existing deep video stabilization methods adopts optical flow-based, which produce artifacts and distortions caused by pixel-level warping and enquire expensive computation time. In this paper, we present a novel unsupervised deep video stabilization approach that addresses the influence of moving objects on video stabilization through robust homography estimation. Specifically, we design a foreground mask estimation module as a preprocessing step using a pre-trained semantic segmentation guided method to distinguish the foreground and background regions, enabling us to estimate camera motion via analyzing the background motion. Additionally, we design a low-level confidence feature extraction module to improve motion alignment loss and ensure robust motion estimation. By integrating the learned low-level confidence features with the foreground mask, we can then design a motion estimation module that captures the consistent spatial correspondence between frames through local and global feature extraction. At last, the learnt robust homography is leveraged to stabilize videos. Our method outperforms related state-of-the-art approaches in both quality and quantity on three public benchmarks while remaining computationally efficient.

The rapid development of information technology such as the Internet of Things, Big Data, artificial intelligence, and blockchain has changed the transaction mode of the financial industry and greatly improved the convenience of financial transactions, but it has also brought about new hidden frauds, which have caused huge losses to the development of Internet and IoT finance. As the size of financial transaction data continues to grow, traditional machine-learning models are increasingly difficult to use for financial fraud detection. Some graph-learning methods have been widely used for Internet financial fraud detection, however, these methods ignore the stronger structural homogeneity and cannot aggregate features for two structurally similar but distant nodes. To address this problem, in this article, we propose a graph-learning algorithm TA-Struc2Vec for Internet financial fraud detection, which can learn topological features and transaction amount features in a financial transaction network graph and represent them as low-dimensional dense vectors, allowing intelligent and efficient classification and prediction by training classifier models. The proposed method can improve the efficiency of Internet financial fraud detection with better Precision,$F1$-score, and AUC.

Recently, network representation learning has been widely used to mine and analyze network characteristics, and it is also applied to blockchain, but most of the embedding methods in blockchain ignore the heterogeneity of network, so it is difficult to accurately describe the characteristics of the transaction. As smart society evolves, Ethereum makes smart contracts reality, while the mine of transaction characteristics appearing on the Ethereum platform is scarce; thus, there is an urgent need to mine Ethereum from contract and transfer. In this article, we propose a heterogeneous network representation learning method to mine implicit information inside Ethereum transactions. Specifically, we construct an Ethereum transaction network by collecting transaction data from normal and phishing Ethereum accounts. Then, we propose a walk strategy that combines timestamps and transaction amounts to represent the information that occurs at the time of a transaction. To mine the types of nodes and edges, we use a heterogeneous network representation learning method to map the transaction network to a low-dimensional space. Finally, we improve the accuracy of the embedding results in the node classification task, which has important implications for Ethereum mining as well as identity recognition.

Abstract Most existing image dehazing methods rely on the solution of the atmospheric scattering model or supervised learning based on paired images. However, owing to incomplete prior knowledge and the lack of paired hazy and haze‐free images of the same scenes as training samples, their performances for single image dehazing are unsatisfactory. Here, the authors present an unpaired image learning method based on the attention mechanism for single image dehazing problems. The method uses the constraint transfer learning ability and circulatory structure of CycleGAN to carry out an unsupervised image dehazing task for unpaired data. Considering the complexity of the haze distribution in actual imaging and human visual characteristics, the improved channel attention and domain attention mechanisms are integrated into the network to process different features and different regions non‐uniformly. The experimental results show that the proposed method achieves good results on both synthetic datasets and real hazy images.

Abstract Spectral clustering is an unsupervised clustering algorithm, and is widely used in the field of pattern recognition and computer vision due to its good clustering performance. However, the traditional spectral clustering algorithm is not suitable for large‐scale data classification, such as hyperspectral remote sensing image, because of its high computational complexity, and it is difficult to characterize the inherent uncertainty of the hyperspectral remote sensing image. This paper uses fuzzy anchors to process hyperspectral image classification and proposes a novel spectral clustering algorithm based on fuzzy similarity measure. The proposed algorithm utilizes the fuzzy similarity measure to obtain the similarity between the data points and the anchors, and then gets the similarity matrix. Finally, spectral clustering is performed on the similarity matrix to compute the classification results. The experimental results on the hyperspectral remote sensing image data sets have demonstrated the effectiveness of the proposed algorithm, and the introduction of fuzzy similarity measure gives rise to a more robust similarity matrix. Compared with existing methods, the proposed algorithm has a better classification result on the hyperspectral remote sensing image, and the kappa coefficient obtained by the proposed algorithm is 2% higher than the traditional algorithms.