Data trusts have emerged as a key governance model in the digital economy, enabling trusted trustees to manage user data securely and privately for value exchange. However, the limited data volume held by individual trustees and strict regulatory barriers to data sharing make complex applications, such as point-of-interest (POI) recommendations, challenging to implement. Furthermore, the inherent sparsity of user behavior records significantly degrades the performance of traditional recommendation systems. To address these challenges, we propose a federated POI recommendation architecture tailored for personal data trusteeship. To address these challenges, we propose a federated POI recommendation architecture tailored for personal data trusteeship. Transcending the conventional client–server paradigm, we design a tripartite framework comprising data owners, data trustees, and an aggregation server, which enables secure, cross-silo knowledge aggregation among institutional trustees. Additionally, we introduce an adaptive multi-factor fusion algorithm driven by semantic-level attention. This algorithm jointly models five key preference dimensions, dynamically weighting heterogeneous features to capture complex user-POI interactions and effectively mitigate data sparsity. Extensive experiments across three datasets demonstrate our system’s exceptional robustness, achieving a 4.00% improvement in Precision@10 over the best baseline.

Accurate emotion recognition of Intelligent Connected Vehicles (ICV) drivers plays a crucial role in mitigating driving risks caused by emotional fluctuations, thereby enhancing road safety and improving driving experiences. Current approaches predominantly suffer from two critical limitations, reliance on single-modality emotional data analysis and neglect of individual personality differences that significantly influence emotional manifestations. To address these challenges, this paper proposes a Personality-aware Multimodal Emotion Recognition Network (PMERNet) that synergistically integrates psychological theory with multimodal learning. Specifically, we first design an end-edge collaborative architecture that optimizes resource allocation between vehicle terminals and roadside units (RSUs), enabling efficient multimodal data collection (end-side) and collaborative feature processing (edge-side). Second, we develop a Weighted Temporal Background Attention module (WTBA-VExt) for robust facial expression analysis and implement a cross-modal self-attention Transformer for explicit multimodal feature fusion. Third, incorporating the psychological Big Five personality theory, we propose a Multimodal Big Five Personality Extractor (Multi-BFPExt) to capture latent personality traits and establish a bidirectional cross-modal fusion mechanism for personality-enhanced emotion recognition. Comprehensive experiments demonstrate that PMERNet achieves statistically significant improvements over existing state-of-the-art methods, reaching superior performance in both accuracy and weighted F1-score. Notably, the weighted F1-score is enhanced by 1.16%.

High mobility and nonindependent and identically distributed (non-iid) data in the Internet of Vehicles (IoV) pose challenges for the convergence speed and accuracy of federated learning model training. Traditional federated learning approaches fail to account for the unique characteristics of local vehicle data, limiting the model performances. To address this, we propose a personalized hierarchical federated learning method tailored for heterogeneous IoV. Our approach customizes models for each vehicle, enhancing their performance on local data. We introduce a network-aware client selection method that dynamically chooses vehicle clients based on their spatiotemporal distribution and changing channel conditions. Additionally, our fine-grained personalized learning method leverages frequency domain analysis to extract and retain both fundamental and personalized knowledge from local model parameters. This allows us to build highly accurate personalized models suited to each vehicle’s data characteristics. Experimental results show that our method improves model accuracy by up to 10.31% and reduces training time by up to 24.93% compared to the latest personalized federated learning algorithms.

Crowdsensed data trading (CDT) solves the problem of data resource scarcity and diversity, faced in conventional data trading by dispatching workers to perform data collection tasks and sharing data through trading. In CDT, both worker and data requesters need to provide geographic location or task location information for spatiotemporal data collection tasks. Existing research has insufficiently addressed the simultaneous consideration of both location privacy information and overlooked the variability in data quality resulting from variations in worker task accessibility and location. To address this problem, we propose a privacy-preserving task allocation scheme with regional coverage based on homomorphic encryption, which allows workers to perform tasks within the qualified region, the degree of regional coverage is associated with data quality to provide diversified data. To solve the sensing data trading and allocation problem for many-to-many users, we further introduce double auction. And thus propose a privacy-preserving data trading scheme to protect bidding information privacy, this scheme ensures the truthfulness of the auction process and mitigates participant manipulation. Besides, we employ a secure multiparty computing strategy to implement truth discovery in CDT, which enables third-party platforms to perform accurate task allocation and winner decisions based on encrypted location and bidding information. Extensive theoretical and simulation analyses show that the proposed scheme satisfies the expected economic properties (truthfulness, individual rationality, etc.), privacy, and effectiveness.

The metaverse is a human-centric beyond-reality virtual world, in which people use virtual identities to live, work, and socialize. Due to the openness and sharing of metaverse applications, the virtual-real identity link (VRIL) may cause uncertainties and unpredictable risks. At present, the research on VRIL risks is still in its infancy and VRIL risk predictions lack a comprehensive theoretical system and methodological tool. In this paper, we first construct a VRIL attack model, according to which an attacker can link a user’s real and virtual identities together using the information observed in the real and virtual worlds. Then we propose the tuple frequency-based VRIL prediction (TupPre) model and discover the population distribution, recursive hypergeometric (RH) distribution, and approximate binomial distribution of the tuple frequency (i.e., the occurrence times of attribute value combinations) given incomplete information. Focusing on the tuple frequency estimation error in biased samples, we introduce attribute value correlation knowledge to improve the prediction performance. The experimental results on generated and real-world datasets show that the TupPre model has excellent performance, with a mean area under the curves (AUCs) of 0.86 to 0.98 on these datasets, and it performs even more superior with certain background knowledge (mean AUC 0.95~0.98). The discovered basic distribution rules of the tuple frequency and the proposed quantitative analysis method for metaverse VRIL risk predictions construct the foundation of the identity privacy framework for the metaverse.

Targeted sentiment analysis aims to identify the sentiment polarity of specific target mentions in a sentence. Existing methods employ neural networks to extract the relations between target mentions and their contexts. Recent approaches based on graph convolutional networks can model the syntactic relations extracted by an external parser into adjacency matrices. However, online reviews are informal and complex, the syntactic structures provided by the parser can be incorrect in these syntax-insensitive scenarios. To remedy this defect, we design a novel integrated syntactic and semantic tree (IS2tree) by labeling semantic relations between the target mention and contexts in a syntactic dependency tree. Furthermore, a dual-channel graph convolutional network (DCGCN) is proposed to encode the contextual information associated with the target mention by dynamic semantic pruning mechanisms and to also retain the syntactic relations. Experimental results demonstrate that the IS2tree has a favorable generalization capability comparing to the state-of-the-art baselines on four public datasets.

Demand-oriented data service can provide the physical entity information for Internet of Things (IoT) users conveniently and quickly. The traditional cloud-oriented data service architecture has been inapplicable to the state time-varying and privacy-sensitive entity data in IoT due to long response delay and the risk of privacy leakage for the entities and users, respectively. The edge-based architecture lacks global service function although it can alleviate problems with cloud services. Moreover, existing demand-oriented data service ignores the characteristics of “thousands of people have thousands of faces” and the implicit intents of users, which results in limited service quality and weak user experience. To solve the above problems, a personalized secure demand-oriented data service scheme is proposed. Specifically, an edge-cloud collaborative architecture is designed to realize privacy-preserving, timely response, and personalized search combining the advantages of edge and cloud. To achieve the personalized service for IoT users, a time span fused personalized ranking method (TSFPR) is proposed to deeply perceive individual demands via mining user preferences with temporal evolution characteristics. Finally, an edge-cloud collaborative personalized secure data service approach (ECPSS) oriented different search modes is presented to achieve encryption data matching and personalized reranking, thereby improving the service quality of the IoT system synthetically. Security analysis and simulation demonstrate the effectiveness of the proposed method in terms of privacy preserving, data service time, and personalized performance.

The metaverse aims to build an immersive virtual reality world to support the daily life, work, and recreation of people. In this survey, the status quo of the metaverse is investigated, and the technical framework of the metaverse is introduced from three aspects: 1) the generation of virtual worlds; 2) the connection of virtual and real objects; and 3) the transmission of data. Specifically, this survey first discusses the development and challenges of the related technologies for virtual world generation methods from three aspects: 1) the 3-D world generation; 2) immersive human–computer interaction experience; and 3) ecosystem. Second, we investigate the status quo of extended reality (XR), motion capture, and brain–computer interface technologies and evaluate the potential and research directions of these entrance technologies for the metaverse. Finally, network and data transmission technologies for the metaverse are reviewed from the Internet of Things (IoT), 5G/6G wireless, and edge computing aspects, the demand side of the metaverse in virtual-reality interpromotion, big data processing, and low-latency networking is discussed, and promising research hotspots are identified.

The high-speed movement of nodes and the burstiness of interactions in the Internet of Vehicles pose huge challenges to the trusted vehicle collaboration and data sharing. Aiming at the disadvantages of existing authentication mechanisms and trust management models for connected vehicles, this article proposes a trust management model enabled by blockchain to ensure the traceability, nontampering, unforgeability, and transparency of vehicular interactions. The proposed trust management model leverages Dirichlet distribution, reputation regression, and revocation punishment to objectively and accurately reflect the trust status of vehicles. Simulation results on real-world data sets show that the proposed trust management model advantageously improves the accuracy of malicious vehicle detection and the attack resistance of connected vehicles.

Federated learning (FL) is a distributed machine learning paradigm that ensures data do not leave local devices. Data sharing problems can be addressed by FL in untrusted environments, e.g., the Internet of Vehicles (IoV). However, FL needs to frequently exchange massive parameters to achieve preset model goals. In addition, the change in bandwidths and the delay of data communications due to user mobility challenge the synchronization of model parameters. In this article, an efficient hierarchical asynchronous FL (EHAFL) algorithm is proposed to adjust the encoding length dynamically according to the bandwidth and reduce the communication cost substantially. A dynamic hierarchical asynchronous aggregation mechanism is proposed leveraging gradient sparsification and asynchronous aggregation techniques to further reduce the communication costs and improve the aggregation efficiency of the global model. Simulation results on MNIST and real-world data sets show that our proposed solution can reduce the communication costs by 98% while only compromising the model accuracy by 1%.

Limited by the storage and computing capacity of Internet of Things (IoT) devices, outsourcing encrypted entity data has become a prevalent trend. The existing IoT entity search methods lack the integration and utilization of both edge and cloud resources and the protection of user privacy. Besides, the traditional searchable encryption mode is inapplicable to state-time-varying entities in IoT. Therefore, in this article, an edge–cloud collaborative entity search method with privacy protection in IoT is proposed, fusing the advantages of edge and cloud resources to fulfill users’ needs for real-time search and privacy protection. Specifically, a secure search architecture and search method for edge–cloud collaboration is designed to support various needs of users, such as real-time search and global search. Then, an adaptive discrimination method for similar interested entities through attribute analysis and feature extraction is proposed to construct attribute-distinguished encrypted index and query vector groups, enabling efficient entity search meanwhile ensuring fast index update. Simulation results demonstrate that the proposed method with privacy protection can effectively improve the efficiency of entity search in IoT while safeguarding user privacy.

Caching AI of Things (AIoT) data at the edge can reduce the load on cloud centers while providing real-time services for AIoT users. Existing static caching strategies based on popularity prediction fail to meet users’ demands for time-varying entity data, while dynamic caching strategies focus only on evaluating the time-varying state characteristics of entity data, but ignore the differences in popularity among entities, resulting in poor service experience. To this end, a device–edge collaborative differentiated data caching strategy considering static entity popularity as well as dynamic system state is proposed. First, a hot entity recognition method centered on user interests is designed to achieve individual preference estimation by mining users’ long short-term interests, and then achieve group interest prediction based on social computing. Based on this, a dynamic caching optimization method is designed, which considers the timeliness of entity data and communication cost of the system to design the objective function of optimal cache decision and then solve it based on reinforcement learning. Simulation results demonstrate that the proposed caching strategy achieves better performance than other benchmark strategies in terms of cache hit rate and search cost.

Wearing masks can effectively inhibit the spread and damage of COVID-19. A device-edge-cloud collaborative recognition architecture is designed in this paper, and our proposed device-edge-cloud collaborative recognition acceleration method can make full use of the geographically widespread computing resources of devices, edge servers, and cloud clusters. First, we establish a hierarchical collaborative occluded face recognition model, including a lightweight occluded face detection module and a feature-enhanced elastic margin face recognition module, to achieve the accurate localization and precise recognition of occluded faces. Second, considering the responsiveness of occluded face detection services, a context-aware acceleration method is devised for collaborative occluded face recognition to minimize the service delay. Experimental results show that compared with state-of-the-art recognition models, the proposed acceleration method leveraging device-edge-cloud collaborations can effectively reduce the recognition delay by 16% while retaining the equivalent recognition accuracy.

The value of trajectory data lies mainly in the spatio-temporal correlation. However, the existing privacy protection methods ignore the spatio-temporal correlation of trajectory data, resulting in a large error in trajectory proportion estimation and Top-K classification. For the privacy of truck trajectory in intelligent logistics, the location and trajectory data perturbation method based on quadtree indexing is proposed, which leverages location generalization and local differential privacy techniques. Our proposed algorithms are suitable for datasets with a large sample space and can protect the trajectory privacy of truck drivers while preserving the strong correlation between adjacent spatio-temporal nodes in the trajectory. The results of simulation on a real trajectory dataset show that the proposed methods not only meet the trajectory privacy requirements of users but also have a good performance in trajectory proportion estimation and Top-K classification.

Targeted sentiment analysis aims to predict the sentiment polarity of the target in a sentence. Most traditional Graph Neural Network-based methods have focused only on the syntactic dependency information of sentences. However, they ignore the position information of words in the linear form of sentences, which leads the model to paying attention to the irrelevant syntactic dependency information to the target. To attenuate the irrelevant information, a novel model called Position-aware Dual Relational Graph Attention Network (PDRGAT) is proposed. Firstly, introducing the position-aware weight window of the syntactic dependency information to make the model pay more attention to the local syntactic information of words that neighbor the target. Secondly, a dual relational attention mechanism combining syntactic and position information is proposed. Experiments show that our model can effectively attenuate the irrelevant syntactic information and outperform state-of-the-art baselines on Accuracy and Macro-F1.

The flexible combination of pervasive portable smart devices and omnipresent high-speed access infrastructures has revolutionized the data sensing and knowledge acquisition in mobile crowdsensing (MCS), underpinning fine-grained city management and highly customizable Internet service applications. However, MCS applications are still confronted with unsolved challenges, such as task assignment, privacy risks, and misbehavior detection. In light of this, this article proposes PETA, a privacy-preserving edge task assignment framework for MCS, leveraging the powerful edge servers deployed between users and the platform to cluster and manage users according to user attributes. Furthermore, group signature is employed by PETA to anonymize and verify user identities for privacy-preserving task assignments. The theoretical analysis and simulation results validate the performance of PETA on identity anonymity, malicious user detection, and task completion rate.

Emerging virtual reality/augmented reality games and self-driving cars necessitate accurate/responsive/private emotion recognition. Usually, traditional emotion recognition models are deployed at central servers, which results in the lack of abilities in generalization and covering the individual variation of clients. This paper proposes a responsive, localized, and private transfer learning based emotion recognition framework under the cloud-edge-client collaborations. Additionally, a 3-dimensional channel mapping method is designed to aggregate features extracted from electroencephalogram (EEG) signals for the generic emotion recognition model, which is further localized and personalized using transfer learning. Simulation results validate the performance of the proposed TLER framework in reducing model training time and improving emotion recognition accuracy.

In mobile crowd sensing (MCS), the cloud as a single sensing platform undertakes a large number of communication tasks, leading to the reduction of sensing task execution efficiency and the risk of loss and leakage of users’ private data. In this paper, we propose a spatial ciphertext aggregation scheme with collaborative verification of fog nodes. Firstly, the cloud and fog collaboration architecture is constructed. Fog nodes are introduced for data validation and slices transmission, reducing computing cost on the sensing platform. Secondly, a multipath transmission method of slice data is proposed, in which the user identity and data are transmitted anonymously by the secret sharing method, and the data integrity is guaranteed by hash chain authentication. Finally, a spatial data aggregation method based on privacy protection is presented. The ciphertext aggregation calculation of the sensing platform is realized through Paillier homomorphic encryption, and the problem of insufficient data coverage in the sensing region is solved by the position-based weight interpolation method. The security analysis demonstrates that the scheme can achieve the expected security goal. The simulation results show the feasibility and effectiveness of the proposed scheme.

How to improve delay‐sensitive traffic throughput is an open issue in vehicular communication networks, where a great number of vehicle to infrastructure (V2I) and vehicle to vehicle (V2V) links coexist. To address this issue, this paper proposes to employ a hybrid deep transfer learning scheme to allocate radio resources. Specifically, the traffic throughput maximization problem is first formulated by considering interchannel interference and statistical delay guarantee. The effective capacity theory is then applied to develop a power allocation scheme on each channel reused by a V2I and a V2V link. Thereafter, a deep transfer learning scheme is proposed to obtain the optimal channel assignment for each V2I and V2V link. Simulation results validate that the proposed scheme provides a close performance guarantee compared to a globally optimal scheme. Besides, the proposed scheme can guarantee lower delay violation probability than the schemes aiming to maximize the channel capacity.