In this study, we address the specific challenges of Named Entity Recognition (NER) in Chinese legal contracts. Traditional Chinese NER systems struggle to balance the loss of important lexical information when using character-based segmentation and the high accuracy and extensive data requirements of word-level segmentation. To bridge this gap, we propose CW-BERT (Character-Word BERT), a novel framework that integrates external lexicon knowledge into BERT through two modules: Character Adapters (CA) and Word Adapters (WA). The Character Adapter enriches each character representation with weighted word-level features using a bilinear attention mechanism, while the Word Adapter updates word representations based on associated character sequences. These adapters are jointly trained with BERT and are injected into intermediate transformer layers, enabling a more expressive and lexically-aware contextual encoding. Extensive comparative experiments on Chinese legal contract datasets demonstrate that CW-BERT achieves better performance in legal contract NER tasks. Furthermore, the dynamic lexicon update mechanism not only proves valuable for enhancing legal knowledge extraction through transferable domain-specific vocabulary, but also demonstrates strong generalization potential in other fields such as medicine-validating its adaptability to diverse domain-specific linguistic features.

Federated learning (FL) provides a privacy-preserving framework for distributed model training across decentralized participating clients (PCs). However, its performance is significantly hindered by data heterogeneity, where non-independent and identically distributed (non-IID) data across PCs induce client drift, compromise model generalization, and may even impede convergence. Although various optimization schemes have been proposed to alleviate these issues, they still suffer from slow convergence and insufficient generalization. In this paper, we propose FedWeight, a novel density-aware adaptive weight optimization in FL that effectively mitigates data heterogeneity without relying on auxiliary data or historical parameters. Specifically, FedWeight allows each client to dynamically and adaptively adjust the loss function weights based on class-wise density derived from kernel density estimation. By assigning larger weights to sparse classes, FedWeight balances learning across sparse and dense classes, thereby accelerating convergence and improving generalization under non-IID settings. Finally, theoretical analysis and extensive experimental results demonstrate that FedWeight achieves more accurate models and significantly faster convergence, improving the performance of integrated baseline approaches.

In Agentic Search, trajectory-level outcome rewards fail to quantify the behavioral contributions of individual steps, while existing step-level reward methods typically rely on costly tree sampling. We view world knowledge as a latent world graph and each IS task as search within a latent task graph, where effective steps should make graph progress toward the answer node. Based on this prior, we propose Graph-Distance Contribution Reward (GDCR), a step-level process reward that scores newly-retrieved and newly-cited entities by their distance to the answer node in a training-time Entity-Relation (ER) graph. We further propose Step Advantage Policy Optimization (SAPO), which converts GDCR into step-level advantages and combines them with trajectory-level outcome advantages. Experiments on four challenging benchmarks validate the effectiveness of our method.

Supervised cross-modal hashing has shown outstanding precision and efficiency in approximate nearest neighbor search (ANNS) on multimodal data. However, the existing methods still face bottlenecks in precision and privacy leakage, which are specifically manifested in the quantization errors of hash codes, insufficient utilization of semantic information, and the lack of inherent adaptability to privacy scenarios. To address these challenges, we propose CrossGuard, a novel privacy-preserving cross-modal retrieval (CMR) framework. It leverages Ternary Quantization Stratified Sampling Sparse Hashing (TARDIS) to achieve high search precision, while employing Distributed Point Function and locality-sensitive hashing to enable private retrieval with sublinear communication complexity. The ternary-sign strategy in TARDIS effectively mitigates the quantization errors in hash codes and optimizes the hash function generated from the least squares method. Furthermore, a stratified sampling strategy using K-means uniformly extracts the semantic information from each label cluster. Moreover, the expressiveness of high-dimensional sparse hashing not only improves the precision, but also makes the CMR endogenously adaptive and scalable for private ANNS. Extensive experiments demonstrate that our framework not only establishes a new state-of-the-art in public retrieval through TARDIS, but also achieves an outstanding balance among privacy, efficiency, and precision with CrossGuard for private retrieval. Specifically, CrossGuard performs a private query in 0.75 ms of private evaluation time on NUS-WIDE with 100,000 entries, while achieving 79.65% image-to-text and 84.15% text-to-image end-to-end precision. See https://github.com/ighxiy/CrossGuard-TARDIS.git for the code.

With the increasing adoption of location-based services (LBSs), managing and protecting spatiotemporal data faces challenges such as inefficient resource utilization, disorganized storage, privacy vulnerabilities, and single points of failure. While blockchain offers decentralized and tamper-proof features, traditional architectures struggle to efficiently store, query, and protect large-scale spatiotemporal data. To address these issues, this paper proposes FST-BlockDAG, a new architecture based on the DAG blockchain. By integrating Voronoi diagrams and the KMeans clustering algorithm, FST-BlockDAG achieves dynamic regional partitioning and optimized storage, enhancing adaptability and processing efficiency in distributed systems. Additionally, FPBFT, an enhanced PBFT consensus mechanism, along with zero-knowledge proofs, ensures Byzantine fault tolerance and data privacy in multinode environments. Using 5G base station data from Xinxiang City, experiments validate FST-BlockDAG’s regional partitioning, consensus mechanism, query efficiency, and privacy protection, demonstrating scalability and security for large-scale data and complex networks. Future work will explore broader applications to enhance adaptability and privacy across different network scales.

Personalized Federated Learning (pFL) is chiefly concerned with improving the capability of the personalized model for each client by integrating global information. In this process, it is crucial that global information can be effectively adjusted to address the unique requirements of individual clients. However, in a heterogeneous environment, significant differences in data distributions across clients can make it difficult for the global model to adapt to all clients. To overcome this challenge, we propose an innovative pFL approach centered on Harmony in Diversity (HD-pFL). HD-pFL uploads the models that rely less on local data, thereby reducing the interference of personalized information during aggregation and improving adaptability of the global model in heterogeneous environments. Additionally, HD-pFL introduces personalized modules which allow clients to further personalize fine-tuning based on shared global models, alleviating challenges caused by heterogeneity and enhancing performance across a broader range of clients. Specifically, at each client, we innovatively utilize the generative model to capture the local attributes, thereby adaptively removing local personalized information through the subtraction of the reconstructed feature vector from the original feature vector. Therefore, the shared model trained on these features further achieves a certain degree of de-personalization. Extensive experiments demonstrate the superiority of HD-pFL in multiple aspects.

Crowdsensing has become a popular method of sensing data collection while facing the problem of protecting participants' location privacy. Existing location-privacy crowdsensing mechanisms focus on static tasks and participants without considering sensing tasks' time requirements and participants' mobility, which cannot achieve satisfactory collected data quality and task completion in crowdsensing with dynamic tasks and participants. Inspired by this, we proposed a location-preservation crowdsensing mechanism, FedSense, considering dynamic tasks and participants based on federated learning (FL) and reinforcement learning (RL). In FedSense, through RL's outstanding decision-making ability, participants select sensing tasks to perform by well-trained RL models without uploading location information to servers for task allocation. We propose an independent tasks selection environment that defines actions, states, and rewards of RL to enable FedSense to achieve satisfactory task completion and data quality while preserving location privacy. Besides, FedSense applies an asynchronous FL aggregation algorithm that reduces participants' network stabilization and device computing ability requirements. Analysis proves that participants' location information does not leave the local device during the model training and task selection process, effectively avoiding privacy leakage. Simulation shows that compared with existing location-preservation crowdsensing mechanisms, FedSense achieves the highest task completion and sensing accuracy for dynamic tasks and participants.

In recent years, contrastive learning has become a prominent approach in unsupervised learning. Based on its powerful feature extraction capabilities, contrastive learning can make full use of the value of a large amount of unlabeled high-quality data for model training, thereby improving model quality. However, contrastive learning still faces serious privacy leakage issues when used in downstream tasks. Although recent research attempted to address this issue by incorporating differential privacy into model training, these approaches still suffer from low model utility, especially in the scenarios of extracting accurate feature by contrastive learning. In this paper, we propose DP-CLMI, a novel methodology by integrating differential privacy into the feature output layer of the contrastive learning model which acts as a feature extractor and is then used by untrustworthy parties for downstream tasks, thereby safeguarding the privacy of the training data. Our experimental results demonstrate that this method not only maintains a robust guarantee of differential privacy, but also preserves acceptable levels of classification accuracy. Additionally, the methodology proves to be resilient against various forms of membership inference attacks, especially neural network-based MIA.

Sidechain techniques improve blockchain scalability and interoperability, providing decentralized exchange and cross-chain collaboration solutions for Internet of Things (IoT) data across various domains. However, current state-of-the-art (SOTA) schemes for IoT multi-domain data exchange are constrained by the need for synchronous networks, hindering efficient cross-chain interactions in discontinuous networks and leading to suboptimal data exchange. In this paper, we propose AsyncSC, a novel asynchronous sidechain construction. It employs a committee to provide Cross-Blockchain as a Service (C-BaaS) for data exchange in multi-domain IoT. To fulfill the need for asynchronous and efficient data exchange, we combine the ideas of aggregate signatures and verifiable delay functions to devise a novel cryptographic primitive called delayed aggregate signature (DAS), which constructs asynchronous cross-chain proofs (ACPs) that ensure the security of cross-chain interactions. To ensure the consistency of asynchronous transactions, we propose a multilevel buffered transaction pool that guarantees the transaction sequencing. We analyze and prove the security of AsyncSC, simulate an asynchronous communication environment, and conduct a comprehensive evaluation. The results show that AsyncSC outperforms SOTA schemes, improving throughput by an average of 1.21 to 3.96 times, reducing transaction latency by 59.76% to 83.61%, and maintaining comparable resource overhead.

Privacy preservation are becoming increasingly significant in machine learning, with recent privacy regulations requiring the deletion of personal data and its impact on models. Although erasing data from storage is simple, removing the influence of data on models remains a challenge. Federated unlearning is an emerging paradigm that aims to forget the knowledge contributed by some specific data to the federated model. In this paper, we design a novel federated unlearning strategy, named FedWiper, which enables exact unlearning in federated learning by erasing specific data and its impact from the federated model. Specifically, based on the granularity of the dataset, we propose training multiple federated submodels to construct a federated unlearning framework, thereby narrowing the scope of the impact of wiped data. Furthermore, the proposed Uni-Adapter structure effectively mitigates the negative impact on model performance from diminishing the dataset scale, while also reducing communication cost. Rather than focusing solely on achieving indistinguishability unlearning of the model for classification task, we extend FedWiper to unlearning for multiple types of tasks and achieve the exact unlearning. Experiments demonstrate that FedWiper can not only accelerate federated unlearning, but also achieve exact unlearning across multiple types of tasks in federated learning while ensuring minimal loss of model performance. Our Code: https://github.com/grey1989/FedWiper.

Bearings are critical in mechanical systems, as their health impacts system reliability. Proactive monitoring and diagnosing of bearing faults can prevent significant safety issues. Among various diagnostic methods that analyze bearing vibration signals, deep learning is notably effective. However, bearings often operate in noisy environments, especially during failures, which poses a challenge to most current deep learning methods that assume noise-free data. Therefore, this paper designs a Multi-Location Multi-Scale Multi-Level Information Attention Activation Network (MLSCA-CW) with excellent performance in different kinds of strong noise environments by combining soft threshold, self-activation, and self-attention mechanisms. The model has enhanced filtering performance and multi-location information fusion ability. Our comparative and ablation experiments demonstrate that the model’s components, including the multi-location and multi-scale vibration extraction module, soft threshold noise filtering module, multi-scale self-activation mechanism, and layer attention mechanism, are highly effective in filtering noise from various locations and extracting multi-dimensional features. The MLSCA-CW model achieves 92.02% accuracy against various strong noise disturbance and outperforms SOTA methods under challenging working conditions in CWRU dataset.

Payment channel hub (PCH), by leveraging a powerful hub to reliably provide off-chain payment services, offers an effective enhancement to payment channel networks (PCNs). However, existing approaches typically rely on a single hub to relay transactions and provide relationship anonymity between participants. This design lacks flexibility under high-frequency transaction scenarios and fail to adequately balance the security of off-chain payments with PCH efficiency. Moreover, current PCNs often adopt source routing, where each transaction path is predetermined without considering the dynamic distribution of large-scale payment requests, leading to load imbalance and even transaction deadlocks. To address these issues, we propose SHARE, a multi-PCH distributed routing scheme based on trusted execution environments (TEE), designed to optimize secure hub allocation and routing efficiency in PCNs. For the multi-hub allocation problem, SHARE balances the management and synchronization costs among participants, and employs mixed-integer linear programming along with supermodular optimization techniques to transform the NP-hard problem into a solvable form, enabling optimal or approximate solutions across various PCN scales. At the routing layer, SHARE integrates global network state with local sender requests to design a TEE-assisted, privacy-preserving distributed routing protocol that dynamically adjusts multipath flow rates, achieving high-throughput and deadlock-free transaction forwarding. We formally prove the security of the SHARE protocol under the universally composable framework. Experimental results demonstrate that SHARE achieves a 43.6

IntroductionTo enhance the quality and yield of strawberries, it is essential to effectively supervise the entire growing process. Currently, the monitoring of strawberry growth primarily relies on manual identification and positioning methods. This approach presents several challenges, including low efficiency, high labor intensity, time consumption, elevated costs, and a lack of standardized monitoring protocols. On the basis of this, there was an urgent need in the market to automate the whole process of target recognition and localization in strawberry growing.MethodsAiming at the above problems, we innovatively constructed a model for target recognition and localization of strawberries based on the YOLOv8s benchmark model, named the WCS-YOLOv8s model. In this paper, the whole growth process of the strawberry was divided into four stages, namely, the bud, flower, fruit under-ripening, and fruit ripening stages, and a total of 1,957 images of these four stages were captured with a binocular depth camera. Using the constructed WCS-YOLOv8s model to process the images, the target recognition and localization of the whole growth process of the strawberry were accomplished. This model proposes a data enhancement strategy based on the Warmup learning rate to stabilize the initial training process. The self- developed SE-MSDWA module is integrated into the backbone network to improve the model’s feature extraction capability while suppressing redundant information, thereby achieving efficient feature extraction. Additionally, the neck network is enhanced by incorporating the CGFM module, which employs a multi-head self-attention mechanism to fuse diverse feature information and improve the network’s feature fusion performance.Results and discussionThe model’s Precision (P), Recall (R), HYPERLINK "mailto:mAP@0.5" mAP@0.5, and mAP@0.5:0.95 of detection were 83.4%, 86.7%, 87.53%, and 60.48%, respectively, and the detection speed was 45.9 FPS(21.8 ms/per image, which significantly improved on the detection accuracy and generalization ability of with the YOLOv8s benchmark model. This model can meet the demand for online real-time target identification and localization of strawberries and provide a new detection method for the automated monitoring and management of the whole growth process of strawberries.

Due to the complex environment of underground coal mines, the quality of captured images is often poor. The dim light in the underground leads to low image contrast and loss of detailed features; meanwhile, the dust and water vapor in the underground makes the image less clear and unevenly bright and dark. In addition, due to the small number of relevant datasets in the underground, it is unable to meet the requirements of deep learning algorithms, which causes great trouble to the underground image enhancement. To address the above problems, this paper proposes an image enhancement algorithm based on CycleGAN network. To address the problem of difficulty in acquiring paired image data in downhole, CycleGAN network is used for label-free training to expand the dataset; to address the problem of loss of image details in downhole, the improved UNET network DCSAU-Net is used instead of the original ResNet network to better preserve the image detail features and structure; by introducing self-attention in the discriminator of CycleGAN mechanism in the discriminator of CycleGAN to further improve the attention and reconstruction ability of important image information; finally, the EM distance with penalty term is used to replace the JS distance to alleviate the difficult problem that the CycleGAN network is not easy to converge. The experimental results show that compared with the original CycleGAN, the algorithm in this paper improves $6.281 \%$, $15.714 \%, 7.627 \%$, and $15.670 \%$ in the four indexes of PSNR, SSIM, IFC, and VIF, respectively. In terms of the visual effect of the images, the improved CycleGAN network generates images with higher clarity and better quality, which proves that the algorithm in this paper can effectively realize the low illumination image enhancement of coal mine underground.

The freshness of aquatic products is directly related to the safety and health of the people. Traditional methods of detecting the freshness of Conger myriaster rely on manual operations, which are labor-intensive, inefficient, and highly subjective. This paper combines computer vision and the DWG-YOLOv8 network model to establish an intelligent classification method for C. myriaster freshness. Through image augmentation, 484 C. myriaster samples were expanded to 2904 samples. The YOLOv8n model was improved by simplifying the network backbone, introducing Ghost convolution and the new DW-GhostConv, thereby reducing the number of parameters and computational load. Test results show that the recognition accuracy of the DWG-YOLOv8 model reached 98.958%, outperforming models such as ResNet18, Mobilenetv3 small, and Swin transformer v2 tiny. The model's parameter count is 16.609 K, the inference time is 57.80 ms, and the model size is only 102 KB. The research provides a reliable method for online intelligent and nondestructive detection of C. myriaster freshness.

The Metaverse, envisioned as the next-generation Internet, will be constructed via twining a practical world in a virtual form, wherein Meterverse service providers (MSPs) are required to collect massive data from Meterverse users (MUs). In this regard, a critical demand exists for MSPs to motivate MUs to contribute computing resources and data while preserving user privacy. Federated learning (FL), as a privacy-preserving collaborative machine learning paradigm, can support distributed intensive computation in the Metaverse. In this work, we first investigate minting the machine learning models into NFT with FL assistance (referred to as FL-NFT), such that MUs as stakeholders can control the ownership and share the economic value of user-generated content (UGC). Specifically, MUs are encouraged to establish a decentralized autonomous organization (i.e., MU-DAO) to aggregate local models and mint FL-NFT. MUs and MSPs optimize the strategies by formulating an imperfect information Stackelberg game to trade off the cost and benefit. We apply the backward induction to derive the equilibrium solution. Then, we construct a privacy-preserving multi-winner sealed-bid auction mechanism (PMS-AM), in which the Hidden Markov Model assists MSPs in choosing rational bidding strategies according to historical bids, and the double auction mechanism determines the winners and price of FL-NFT. Finally, the numerical results based on theoretical analysis and simulations demonstrate that the proposed PMS-AM can increase the quality of FL-NFT and achieve the economic properties of incentive mechanisms such as individual rationality and incentive compatibility.

Robotically selective broccoli harvesting requires precise lightweight detection models to efficiently detect broccoli heads. Therefore, this study introduces a lightweight and high-precision detection model named LBDC-YOLO (Lightweight Broccoli Detection in Complex Environment—You Look Only Once), based on the improved YOLOv8 (You Look Only Once, Version 8). The model incorporates the Slim-neck design paradigm based on GSConv to reduce computational complexity. Furthermore, Triplet Attention is integrated into the backbone network to capture cross-dimensional interactions between spatial and channel dimensions, enhancing the model’s feature extraction capability under multiple interfering factors. The original neck network structure is replaced with a BiFPN (Bidirectional Feature Pyramid Network), optimizing the cross-layer connection structure, and employing weighted fusion methods for better integration of multi-scale features. The model undergoes training and testing on a dataset constructed in real field conditions, featuring broccoli images under various influencing factors. Experimental results demonstrate that LBDC-YOLO achieves an average detection accuracy of 94.44% for broccoli. Compared to the original YOLOv8n, LBDC-YOLO achieves a 32.1% reduction in computational complexity, a 47.8% decrease in parameters, a 44.4% reduction in model size, and a 0.47 percentage point accuracy improvement. When compared to models such as YOLOv5n, YOLOv5s, and YOLOv7-tiny, LBDC-YOLO exhibits higher detection accuracy and lower computational complexity, presenting clear advantages for broccoli detection tasks in complex field environments. The results of this study provide an accurate and lightweight method for the detection of broccoli heads in complex field environments. This work aims to inspire further research in precision agriculture and to advance knowledge in model-assisted agricultural practices.

The pH and firmness index (FI) of red globe grapes determine the taste and post-harvest quality of the fruit. In this paper, a method for the detection of pH and firmness of red globe grapes based on transmission spectroscopy technology was proposed. Spectral datas were first collected from 360 full-growth-cycle red globe grape samples, which were pre-processed and modeled by different spectral pre-processing methods as a way to determine the best spectral pre-processing method. Then competitive adaptive reweighted sampling (CARS), successive projection algorithm (SPA), uniformative variable elimination (UVE) and CARS-SPA, UVE-SPA composite data dimensionality reduction methods were used respectively for extracting feature variables from spectra. Finally, partial least squares regression (PLSR) detection models for pH and firmness of red globe grapes were established, respectively. The optimal prediction models for pH and firmness of red globe grape samples were moving-average method (MA)-CARS-SPA-PLSR and MA-UVE-SPA-PLSR. The correlation coefficient of prediction (RP) of the prediction sets of the two models were 0.9882 and 0.9588, and the residual predictive deviation (RPD) were 6.5857 and 3.5167, respectively. The results showed that transmission spectroscopy could be applied to the detection of pH and firmness of red globe grapes, which provided a new idea and a new method for the detection of pH and firmness of red globe grapes in the whole growth cycle.

Soluble solids content (SSC) and pH of red globe grapes are crucial measures of quality. In this paper, we used hyperspectral imaging technology to achieve nondestructive detection and distribution visualization of SSC and pH of red globe grapes. First, the hyperspectral images of samples were collected. Then, CARS, SPA, GA, IRIV were used to extract feature variables from raw spectral (RAW) information. The PLSR prediction models of samples were developed. By comparing the different prediction models, RAW-IRIV-PLSR was selected as the optimal model. Finally, the SSC and pH of the samples were calculated to obtain a grayscale image and perform a pseudo-color transformation to visualize the distribution of SSC and pH. By studying the classification of the maturity of samples, it was concluded that the best discriminant classification model of maturity was RAW-IRIV-ELM. Hyperspectral also provided a new method for maturity stage classification of red globe grapes.