Multi-access edge computing (MEC) is an important technology to accelerate the response speed of computation-intensive and delay-sensitive tasks, which promotes the development of Artificial Intelligence (AI) and Internet of Things (IoT). Federated learning (FL) over mobile devices, coupled with MEC to build an intelligent network, can avoid the risk of privacy leakage by keeping the sensitive information of each agent locally. However, there exist some problems in implementing FL over mobile devices, such as the shortage of edge bandwidth and computing resources. In addition, network dynamics and the heterogeneity of mobile devices need to be considered. To address these issues, we propose a heterogeneous Stackelberg game approach based on deep reinforcement learning (DRL) to achieve the desired trade-off between computing and communication in the FL system, called Energy Efficient Heterogeneous Federated Learning (EEHFL). Specifically, EEHFL designs a new two-stage Stackelberg game approach based on the heterogeneous FL architecture with convergence guarantee targeting efficient energy, which is modeled separately. Furthermore, DRL algorithms are introduced to solve the problem, realizing the control of heterogeneous parameters in a dynamic environment. Experimental results illustrate that compared with state-of-the-art baselines, our model achieves remarkable improvement, which demonstrates the superiority of our model on saving cost and energy consumption.

Existing keyphrase generation (KG) methods typically employ a weight subnetwork to combine the copy and generating probability distributions when producing target keyphrases. However, these methods are hampered by irrational copy and generating distribution proportions from the weight subnetwork. To tackle this problem, we propose a category guided meta-tuning keyphrase generation method called CMeta-KG, which utilizes category information and alternate meta-tuning to optimize the weight subnetwork more accurately. With contrastive learning, CMeta-KG adopts a retriever to mine richer category information and yields more accurate distribution proportions. Moreover, CMeta-KG revises the standard meta-learning pipeline with an alternate parameter update strategy to refine the weight subnetwork. Notably, we provide a theoretical analysis based on information theory to explain the choice of focusing on the meta-tuning weight subnetwork instead of the overall model parameters. Experiments on the scientific article and social media datasets demonstrate the effectiveness of the contrastive setup and meta-tuning pipeline.

Vehicular Ad Hoc Networks (VANETs) are critical for enhancing road safety and optimizing traffic flow. Attribute-based searchable encryption (ABSE) with blockchain supports one-to-many encryption, ciphertext search, and decentralized storage, providing a solution to enable data sharing in VANETs. However, existing VANETs schemes face many limitations: blockchain miner nodes are at risk of stealing vehicle attribute keys during key verification; the lack of permission control may also lead to unauthorized downloads; and multi-keyword search, encryption, and decryption require a lot of computing power. To address these issues, we propose an attribute-based searchable encryption scheme for secure data sharing with blockchain (ABSE-VANETs). Specifically, we introduce multiple authorization authorities to manage attributes and generate obfuscated key parameters via bilinear pairing to prevent the leakage of vehicle attribute keys. We combine linear integer secret sharing (LISS) to bind the ciphertext download permission and the vehicle attribute set to prevent unauthorized ciphertext download. Moreover, we optimize the multi-keyword index generation algorithm and trapdoor structure and construct a multi-keyword index with the help of blockchain, which achieves the fast positioning of the ciphertext address on the chain. Finally, we utilize the edge cloud server to realize proxy decryption of ciphertexts, which further reduces the computation pressure on the vehicle terminal. Formal security analysis and performance comparison experiments show that our scheme is highly secure under the DBDH assumption and performs well in terms of communication and computation efficiency.

Sequential recommendation predicts users’ potential behaviors based on their historical behavior sequences. In recent years, most of the research on sequential recommendation is based on the self-attention mechanism. However, this mechanism focuses on low-frequency signals and is insufficient in modeling users’ short-term interests. Moreover, there is noise and uncertainty in users’ behavior sequences, making it difficult for a single-vector representation to capture complex situations. To address these issues, this paper proposes the GFASRec model, which integrates Gaussian distribution modeling and adaptive frequency filtering. Gaussian embeddings are used to represent items and side information as distributions, where the mean represents stable preferences and the covariance represents uncertainty. We also design a Multi-scale Global Filter, which captures temporal patterns across different scales by adjusting weights based on frequency. Additionally, we introduce a Shared-Routed Mixture of Experts to balance global and dynamic modeling and improve representational diversity. Experiments on multiple real-world datasets show that GFASRec outperforms advanced baseline models, verifying its advantages in robustness and expressive power.

The interconnection of electronic health devices through the Internet of Things (IoT) technology plays a crucial role in improving healthcare services. The existing schemes usually use traditional cryptographic algorithms and secret sharing to protect the privacy of user health data records generated by smart electronic devices. However, key transmission is exposed to the risk of leakage, and it is vulnerable to quantum attacks. To address these issues, this article proposes a novel data privacy protection named lattice-based threshold encryption for data privacy protection with zero trust in Internet of Medical Things (LTEN-ZTIoMT). Specifically, this article proposes a lattice-based encryption algorithm by introducing zero trust and threshold secret sharing in managing users and IoT devices, distributing private keys, thus enhancing the ability to resist quantum attacks in LTEN-ZTIoMT. To prevent tampering in transmitting private key subshares, this article adopts the oblivious transfer protocol, where the private key subshares escrow only receives his selected private key subshares without gaining other private key subshares information. Moreover, we design a dynamic authentication and authorization mechanism based on a whitelist, which enables the dynamic verification of IoT devices and the revocation of certificates. To prevent unauthorized users from accessing the private key subshares and data ciphertexts, we have introduced the token verification mechanism, in which the user identity can be verified by the private key subshares escrow. The formal security analysis in the Scyther tool and the theory analysis of the proposed scheme are provided. The thorough performance simulation experiments on private key share distribution, private key recovery, encryption, decryption, throughput, and energy consumption are conducted, demonstrating the effectiveness of the proposed LTEN-ZTIoMT.

Vehicle-to-Grid (V2G) technology enables electric vehicles to obtain power from the smart grid and supply electricity back during peak hours, and it is crucial for enhancing energy efficiency and reducing carbon emissions. However, attackers may steal sensitive information from electric vehicles through frequent charging and discharging transactions, or even forge such transactions to breach V2G security. Existing schemes mainly adopt ring signatures to achieve forgery resistance and privacy preservation for charging and discharging transactions, and extend them into multi-layer ring signature structures to support multi-input and multi-output transactions. Nevertheless, such schemes either cause the signature size and verification complexity to grow with the transaction scale, or rely on logarithmic compression proofs that introduce substantial system overhead, thereby limiting their practical applicability. To address this issue, we propose an efficient blockchain-enabled linkable ring signature scheme named V2G-RingChain. Specifically, we design an innovative multi-layer linkable ring signature protocol, which reconstructs the key images structure to reduce the computation time required for signing and combines Sum Arguments of Knowledge to control the signature size at a logarithmic level. Moreover, we construct a blockchain-enabled ring transaction mechanism to decouple anonymous electricity payments, blockchain verification, and bill extraction, which relies on Pedersen Commitments and Back-Maxwell RangeProof for hidden account amounts and non-negative judgment. Finally, we conduct a comprehensive security analysis demonstrating that our scheme ensures security and privacy. We conducted extensive experiments, and the results show that V2G-RingChain significantly reduces signature generation cost, verification cost, and communication cost compared with state-of-the-art schemes.

The task of table evidence retrieval of Web Information aims to identify the most relevant Web tables from an extensive corpus for supporting Web information retrieval and text analytics. However, previous methods suffer from low accuracy in scenarios with limited computing resources and incapability of assessing the authenticity of statements with sufficient evidence, posing potential risks to Cybersecurity and spread of false information. To address these challenges, we first restructure three existing table fact verification datasets for evaluating the performance of evidence retrieval of Chinese Web Tabular Data, which accounts for a large proportion in the Web. And then, we propose a novel table evidence retrieval model based on Dense Passage Retriever for enhancing retrieval accuracy in scenarios with limited computational resources. Our experimental results across multiple datasets demonstrate that the proposed model achieves commendable accuracy, validating its effectiveness.

Text representation models, such as Transformers and RNNs, are foundational to natural language processing research and play a crucial role in a wide range of downstream tasks. However, most existing large models rely on the self-attention mechanism, which involves frequent token interactions, a large parameter scale, and substantial hardware and data resource requirements-posing significant challenges for deployment on mobile and edge devices with limited computing power, high communication latency, or intermittent connectivity. To address these practical constraints, this study proposes an ultra-lightweight text representation model based on Causal Recurrent Convolutional Networks (CausalRCN), specifically engineered for efficient inference in resource-constrained mobile environments. Rather than introducing new atomic modules, our innovation lies in the systematic integration of causal convolution and self-enhancement mechanisms into a compact recurrent architecture that eliminates attention-driven computation entirely. By replacing self-attention with parallelized causal convolutions and recurrent feature propagation, the model achieves dramatically lower computational complexity and memory footprint, enabling real-time inference on commodity edge hardware. The design leverages local causality to approximate global contextual dependencies and employs feature self-enhancement to strengthen nonlinear expressiveness, effectively balancing accuracy and efficiency. Experimental results show that with only 666 K parameters, the proposed model achieves over 10 & times; higher throughput and 50% less memory usage compared to competitive baselines, while improving accuracy by more than 1%. Validated on real-world platforms including Raspberry Pi 4B and Qualcomm Snapdragon Gen 1, the system demonstrates up to a 90 & times; speedup over ALBERT in end-to-end latency, underscoring its suitability for mobile intelligent computing.

Large Language Models (LLMs) have demonstrated strong capabilities in knowledge-based Visual Question Answering (VQA). However, existing prompt construction methods are often rigid and fail to fully exploit the reasoning potential of LLMs. To address this limitation, we propose the Intelligent Prompt Construction Framework (IPCF), which equips an autonomous agent with the ability to dynamically generate task-specific prompts. IPCF consists of a planner and a toolbox: the planner, powered by an LLM, enables autonomous decision-making, while the toolbox provides three tools—the vanilla VQA model for inspiration, the LLM for knowledge injection, and a knowledge base for information retrieval. This architecture allows the agent to flexibly determine when and how to invoke each tool and to construct adaptive prompts accordingly. Experimental results show that IPCF achieves performance gains of 2.6 and 1.9 points over existing baselines on the OK-VQA and A-OKVQA datasets, respectively.

Aspect-based sentiment analysis task aims to predict the sentiment polarity of specific target aspects in sentences, which is critical for fine-grained opinion mining but faces challenges such as ambiguous sentiment boundaries and insufficient capture of structural-semantic correlations. Existing methods often struggle to balance discriminative feature learning and multi-dimensional information integration, leading to limited performance in complex scenarios. To address these issues, we propose a Ternary Paradigm-Oriented and Hybrid-Graph Capsule Aggregation (TPHCA) model. Specifically, we design a ternary paradigm construction module, which enhances the model's ability to distinguish between different sentiment polarities by building triplet combinations of positive examples, anchor sentences, and negative examples. Additionally, we construct a hybrid graph module that includes a syntactic graph and a semantic graph to provides comprehensive structural-semantic representations. The syntactic graph models explicit grammatical dependencies based on hierarchical weights, while the semantic graph captures implicit semantic correlations via cosine similarity. A capsule-enhanced graph neural network aggregator is further proposed to hierarchically aggregate dual-graph features through multi-round graph convolution and dynamic routing mechanisms, effectively extracting aspect-specific sentiment features. Experimental results demonstrate that the TPHCA model significantly outperforms existing methods on multiple benchmark datasets, with F1 score improvements of 6.38%, 1.90% and 9.98% on the Laptop, Restaurants and Twitter datasets, respectively. Moreover, ablation experiments validate the effectiveness of each component of the model.

In latency-sensitive IoT Consumer Electronics (CE) scenarios, effective task offloading is pivotal for ensuring interaction quality. However, existing approaches often overlook fine-grained task urgency, while standard Deep Reinforcement Learning (DRL) suffers from slow convergence and high inference latency. To address these challenges, this paper proposes a semantic-aware optimization framework grounded in the Holographic Counterpart (HC) model. First, the Historical-Experience-Driven Dynamic Priority Task Allocation and Offloading (HE-DPTAO) method is developed, which leverages semantic priorities and edge caching to minimize queuing delays and task drop rates. Subsequently, the Offloading Decision Method based on MDPRL (ODM-MDPRL) is introduced, integrating self-attention mechanisms and asynchronous prioritized experience replay to accelerate training efficiency and decision precision. Simulation results demonstrate that this coordinated framework significantly outperforms representative baselines, achieving a superior balance between timeliness, stability, and resource utilization.

The protection of data privacy in Internet of Things (IoT)-enabled federated recommendation systems (FRSs) through federated learning plays a crucial role in preventing the leakage of raw user data generated by massive interconnected devices. However, most existing federated recommendation methods mainly employ model aggregation to achieve collaborative training, while lacking robust mechanisms for identity privacy protection, which leaves user identities vulnerable to source inference and membership inference attacks. To address these issues, we propose a privacy-preserving scheme named ring signatures for privacy protection with hybrid encryption in IoT-enabled federated recommendation systems (RSHE-IoTFRS). Specifically, we propose a linkable ring signature mechanism that hides individual user identities within a group and generates a link label to prevent user identification, thereby ensuring secure and auditable participation in federated recommendation in IoT-enabled environments. Moreover, we employ a hierarchical encryption strategy to construct a secure parameter transmission framework, which guarantees the end-to-end confidentiality of model updates. We design a selective aggregation (SA) method that integrates item clustering and user similarity to alleviate data heterogeneity during collaborative training. Finally, we conduct experiments on five real-world datasets, and the results demonstrate that the proposed scheme outperforms state-of-the-art schemes on average in hit rate (HR@10) and normalized discounted cumulative gain (NDCG@10), while signature verification and encryption decryption operations are completed within milliseconds.

Click-through rate (CTR) prediction is a core task in recommender systems, aiming to estimate the probability that a user clicks a candidate item. With rich content availability, multimodal signals such as text and images have been incorporated to complement ID-based features. However, users often focus on different modalities depending on interests and context, making simple concatenation or static fusion insufficient to capture dynamic preferences. To address this, we propose an Adaptive Preference Fusion (APF) framework for multimodal CTR prediction. Multimodal representations are projected into a shared semantic space and modeled with LightGCN to capture user–item interactions. A Dynamic Attention Fusion (DAF) module adaptively adjusts modality weights according to their contribution to CTR prediction. A cross-modal correlation constraint further enhances dominant representations and improves robustness. Experiments on three benchmark datasets results indicate that APF performs better than current leading methods on both AUC and LogLoss.

Driven by distributed computing, blockchain and large language model, etc., computational tasks are gradually shifting from centralized simple tasks to decentralized complex tasks. In a decentralized task scenario, all the participants are on an equal footing. Determining whether the participating parties are trustworthy, and reaching a credible consensus in situations where untrustworthy parties exist, has always been a tricky issue. To address this issue, we utilize the reliability of the network to enhance the RAFT algorithm and put forward a novel consensus algorithm NR-RAFT. In NR-RAFT, we integrate node reliability and link reliability to quantitatively evaluate the reliability of consensus results, which is called consensus reliability. Specifically, the reliability of nodes depends on their reputation and the reliability of links that are connected to them. We achieve the evaluation by running an improved ET algorithm and a packet statistics-based approach, respectively. Furthermore, we introduce a “maximum reliability election principle” to avoid the conflicts in the process of leader election. To improve the reliability and efficiency of consensus, we select the highly trustworthy members in the network to form the consensus committee and introduce a salp swarm algorithm-based member optimization method to provide flexibility in the composition of the committee and protection against targeted attacks by malicious nodes. Finally, the experimental results prove the effectiveness of NR-RAFT.

Digital human generation models play a crucial role in industrial applications, offering innovative directions to reduce costs and enhance benefits across fields such as education, entertainment, and marketing. However, most existing methods focus on isolated modalities within multi-modal processes, unable to synchronize semantic information with the required consistency and interdependency through prompts generated for each modality, resulting in semantic misalignments in visual appearance, auditory timbre, and emotional expressions. To address these issues, we propose VividTalker, a unified multimodal approach for text-to-digital human generation that prioritizes semantic consistency through a fully automated pipeline encompassing the synthesis of text-to-image, text-to-audio, and (image & audio)-to-video. Specifically, VividTalker generates consistent and interdependent prompts for each modality during the pre-generation stage. Then, it utilizes an Artificial Intelligence (AI) agent driven by a Small Language Model (SLM) to dynamically coordinate visual representation, auditory features, and semantic logic, ensuring improved cross-modal semantic alignment in the generated videos. Since the SLM-driven AI agent requires lower computational power and offers higher throughput, Vividtalker achieves faster generation compared to the Large Language Model (LLM) driven one. Moreover, to assess the effectiveness and stability of VividTalker, we incorporate the task success rate as one of the evaluation metrics, reducing the influence of subjective evaluation. Experimental results indicate that VividTalker achieves an 87.76% similarity to real human videos while improving the generation speed by 18.49%. This strikes a well-balanced compromise between generation speed and quality while delivering exceptional cross-modal consistency.

Existing Data Augmentation (DA) methods face key limitations: traditional approaches like Easy Data Augmentation (EDA) may introduce semantic distortions, while Pretrained Language Models (PLMs) often generate insufficiently diverse data. To address these challenges, we propose a Large Language Model (LLM) based framework that integrates EDA, PLM-based augmentation and LLM-based generation to produce the augmented data. Our method incorporates a reranking mechanism to filter generated samples, ensuring semantic consistency while maintaining diversity. Additionally, we utilize EDA and PLM-generated data as prompts to guide the LLM, leveraging its capabilities to further enrich the augmented data. Experiments demonstrate that our framework significantly improves the performance, achieving accuracy gains of 3.48 and 0.50 on the STSA and SNIPS datasets.

Graph Neural Networks (GNN) possess powerful relational modeling capabilities, making them a leading solution for capturing and understanding users’ latent preferences, thereby significantly advancing personalized recommendation systems. However, current GNN-based recommendation methods heavily depend on explicit and static data, which results in overlooking the semantic value of user and item information. This oversight can lead to biases in understanding user preferences thus degrade recommendation performance. To address this problem, we propose a novel Dual-domain Contrastive Reinforcement Large Language Model for Recommendation (DCRLRec), which leverages large language models (LLMs) to perform inference across both the textual and graph domains, while applying contrastive reinforcement to enhance the alignment and representation of user and item nodes for personalized recommendations. The DCRLRec includes three key modules: collaborative domain feature perception module, semantic graph domain reinforcement module and contrastive alignment module. Dual-domain information leverages the advanced reasoning capabilities of the LLM to augment the user interaction features of the collaborative domain and the semantic graph domain to capture complex semantic and structural information about items to produce different representations for users and items, respectively. Furthermore, a cross-domain contrastive reinforcement method is introduced to align embeddings from both domains, ensuring high-quality user recommendations. Through experiments on two benchmark datasets, compared with the state-of-the-art baselines, the extensive results exhibit that DCRLRec achieves competitive improvements of up to 3.61% in AUC and 2.21% in F1 scores, respectively.

The rapid growth of Internet of Things (IoT) boosts device connectivity but complicates cross-domain interoperability. Centralized authentication faces single-point failures, while blockchain’s decentralized, tamper-resistant nature overcomes these limits. Current work optimizes blockchain-based cross-domain authentication efficiency and user privacy, yet IoT’s dynamic environments require tailored mechanisms. To tackle these issues, we designed a decentralized cross-domain authentication framework, DynaTrac, with dynamic reputation weighting and accountable tracing mechanisms. DynaTrac uses Distributed Key Generation (DKG) and Aggregate Threshold Signatures (ATS) protocols to generate shared keys and signatures, avoiding the involvement of trusted third parties. In order to adapt the mobility of the device, we design a dynamic reputation-weighted DKG-ATS protocol to cope with the frequent entry and exit of IoT devices. This protocol dynamically adjusts the reputation and weights of the nodes based on behavioral metrics. The updated reputation weights govern nodes’ participation shares in key negotiation and authentication processes, effectively minimizing disruptions caused by low-weight devices. The framework employs authentication tokens, which are inspired by blockchain transaction structures, to enable efficient cross-domain verification. Additionally, a signature-based traceability mechanism embeds traceable tags into threshold signatures to ensure node anonymity during normal operations while enabling the identification of malicious behaviors. Finally, security analysis and simulation experiments validate the effectiveness of DynaTrac.

The task of recognizing Chinese variant characters aims to address the challenges of semantic ambiguity and confusion, which potentially cause risks to the security of Web content and complicate the governance of sensitive words. Most existing approaches predominantly prioritize the acquisition of contextual knowledge from Chinese corpora and vocabularies during pretraining, often overlooking the inherent phonological and morphological characteristics of the Chinese language. To address these issues, we propose a shared-weight multimodal translation model (SMTM) based on multimodal information of Chinese characters, which integrates the phonology of Pinyin and the morphology of fonts into each Chinese character token to learn the deeper semantics of variant text. Specifically, we encode the Pinyin features of Chinese characters using the embedding layer, and the font features of Chinese characters are extracted based on convolutional neural networks directly. Considering the multimodal similarity between the source and target sentences of the Chinese variant-character-recognition task, we design the shared-weight embedding mechanism to generate target sentences using the heuristic information from the source sentences in the training process. The simulation results show that our proposed SMTM achieves remarkable performance of 89.550% and 79.480% on bilingual evaluation understudy (BLEU) and F1 metrics respectively, with significant improvement compared with state-of-the-art baseline models.

Multi-dimensional spectrum prediction is essential for spectrum sharing and dynamic spectrum access (DSA), tack-ling spectrum scarcity and improving wireless communication. Traditional methods often use machine learning (ML), which requires manual feature extraction, or deep learning (DL), which demands high computational resources. This paper proposes a lightweight multi-dimensional spectrum prediction model using an adaptive broad learning network (ABLN). The model employs a sliding window to preprocess data and establishes input layers using randomly generated feature and enhancement nodes. The weights of broad learning are determined by solving the pseudo-inverse, and the structure is incrementally extended without retraining, reducing computational complexity. An adaptive node increment module optimizes hyperparameters efficiently. Experimental results demonstrate that ABLN reduces computational overhead while maintaining robust prediction performance across various scenarios.