最新刊期
-
Vulnerability Detection Based on Graph Neural Networks： A Global Counterfactual Explanation Method AI导读
摘要：Graph Neural Networks (GNNs), renowned for their exceptional ability to capture the semantic structure of source code, have become a vital tool for code feature learning in software vulnerability detection. However, the black-box nature of GNNs severely hampers their application in the vulnerability detection domain. To address this challenge, existing research has proposed explainers based on counterfactual reasoning, which provide causal explanations by identifying minimal perturbations to the code graph that alter the model's prediction. Nevertheless, these methods generally lack guidance from global information, and their perturbation scope is confined exclusively to the edges of the code graph. To tackle this issue, we propose GlobalGCE, a novel counterfactual reasoning-based explainer designed to enhance the predictive explainability of GNNs-based vulnerability detection models. GlobalGCE generates counterfactual explanations by identifying the minimal set of edge and node perturbations in the code graph. It further integrates prior knowledge from a global corpus of vulnerable program graphs to accelerate the search for the counterfactual explanation subgraph. Experimental results on four mainstream GNNs-based vulnerability detection models demonstrate that GlobalGCE significantly outperforms current state-of-the-art factual and counterfactual reasoning explainers.关键词：Graph Neural Networks(GNNs);vulnerability detection;counterfactual explanation;global explanation;model explainability;code graph;graph perturbation3|0|0更新时间：2026-10-02 -
Multi-Server Bidirectional Verifiable Computation Protocol Based on Batch Polynomial Commitment AI导读
摘要：In outsourced computation, clients typically delegate computational tasks to multiple servers. In a security model where the client and server do not trust each other, this paper addresses the issues of computational integrity, input privacy, and input validity in multi-server outsourced computation, and proposes BatchProofs, a multi-server bidirectional verifiable computation protocol based on batch polynomial commitments. To meet the verification requirements of clients and servers in the protocol, two polynomial commitment-based verification schemes are designed: a one-to-one scheme and a one-to-many scheme. The one-to-one verification scheme achieves local aggregation of proofs for bivariate polynomials and is used by clients to verify the correctness of computation results. The one-to-many verification scheme efficiently generates proofs in batches and aggregates them through random linear combinations, and is used by servers to verify the validity of input secret shares. Experimental results show that the proof verification time of the one-to-one verification scheme is about 10 times faster than that of the Kate-Zaverucha-Goldberg (KZG) scheme. When the number of certificates generated in bulk is -
- 摘要：With the widespread adoption of blockchain airdrop activities, sybil attacks have emerged as a significant threat to the security of the Web3 ecosystem. By controlling large numbers of pseudonymous identities, attackers obtain improper benefits and severely undermine the fairness of decentralized applications. To detect and mitigate such attacks, various on-chain sybil address detection methods have been developed. However, existing methods still suffer from insufficient node feature extraction, limited temporal modeling capability, and strong dependence on manually labeled data. To address these challenges, this study proposes an on-chain sybil address detection model based on temporal dynamic graph neural networks. The proposed model first constructs a time-aware dynamic directed graph for invitation relations and extracts 68-dimensional node features across four dimensions: topological structure, behavioral statistics, social interactions, and network centrality. A temporal dynamic graph representation learning algorithm that integrates a memory mechanism with a temporal graph attention mechanism is then proposed, where the memory mechanism is responsible for maintaining long-term node states and the temporal graph attention mechanism is used to aggregate temporal information from neighboring nodes. Finally, an asymmetric contrastive learning framework is introduced to perform self-supervised training, effectively reducing reliance on manually labeled data. Experiments conducted on a real-world blockchain invitation dataset demonstrate that the proposed model achieves an F1 score of 91.11%, an AUC of 97.69%, and a balanced accuracy of 92.50%, outperforming the baseline methods across all metrics.关键词：sybil address detection;node feature extraction;temporal dynamic graph neural networks;contrastive learning21|50|0更新时间：2026-09-23
-
- 摘要：Field-effect transistor (FET) biosensors offer advantages such as high sensitivity, label-free detection, real-time monitoring, and ease of miniaturization and integration, rendering them highly valuable for the detection of extracellular vesicles (EVs) as disease biomarkers. This review first outlines the biological characteristics and clinical diagnostic significance of EVs, and elucidates the fundamental detection principles of FET biosensors. On this basis, the design strategies of FET biosensors are systematically summarized. Subsequently, with channel materials as the main thread, the latest research advances in EVs detection using FET biosensors based on graphene, carbon nanotubes, metal oxides, organic electrochemical transistors, and other types are critically reviewed. Finally, key challenges regarding improvement of specificity, interference in complex sample matrices, and clinical translation are discussed, and future trends toward integration, standardization, and point-of-care testing are prospected. This review aims to provide a clear and comprehensive framework for the diverse pathways of FET-based platforms in EVs detection, and to offer a theoretical reference for promoting fundamental innovation and clinical translation in this field.关键词：field-effect transistor (FET);biosensor;biomarker;extracellular vesicles (EVs);liquid biopsy32|51|0更新时间：2026-09-18
- 摘要：Head pose estimation, a key research task in computer vision, has broad application prospects in human-computer interaction, security monitoring and other scenarios. However, practical applications face many challenges, such as severe occlusion, insufficient lighting and extreme posture angles. Therefore, this article proposed a head pose estimation method based on the representation of rotational displacement relationships of facial features, which mainly included two modules: Facial Relationship Mining (FRM) Module and Covert Displacement Feature Representation (CDF) Module. Specifically, the FRM Module significantly enhanced the model’s ability to perceive and understand key head pose information by finely modeling the spatial distribution and interdependencies of facial regions. The CDF Module constructed a robust feature structure suitable for occluded scenarios centered on the relative displacement of facial features during rotation, effectively breaking through occlusion interference. Experimental results on BIWI, AFLW2000 datasets and their occluded versions show that the proposed method outperforms several state-of-the-art methods in occluded conditions while maintaining comparable accuracy in non-occluded scenarios, fully verifying its effectiveness and robustness.关键词：head pose estimation;facial relationship mining;covert displacement feature;Transformer;image understanding5|31|0更新时间：2026-09-17
- 摘要：As a key indicator of logistics service levels, leading time has a significant impact on consumers’ expected utility and the pricing of logistics services.It is necessary to research on the mechanism of interaction between above three factors. A service supply chain system comprising a logistics platform provider, a self-operated logistics provider (SLP), and a franchise logistics provider (FLP) is established to deliver logistics services to consumers. First, the consumer expected utility function and demand function are characterized based on the interaction between leading time and logistics service pricing. Subsequently, the logistics service pricing strategies are studied under two scenarios, i.e. identical lead time (IDL) and differentiation lead time (DLT), and divided into the growth stage (the first stage) and maturity stage (the second stage) .. The research findings are as follows: 1) The consumer expected utility function, which comprehensively considers leading time and logistics service pricing, is related to the development stages and the channels through which logistics service provider.; 2) In the growth stage, the optimal pricing and costs of the SLP and the FLP are negatively correlated with leading time. The optimal pricing strategies should be formulated under the condition where the consumer expected utility is balanced; 3) In the maturity stage, there is a threshold value for leading time, based on which optimal pricing strategies can be formulated for both the SLP and FLP. The SLP can significantly shorten leading time with limited price reductions, while the FLP will adopt a follow-the-leader pricing strategy. The research conclusions provide a reference for enterprises to make business decisions such as pricing and leading time in logistics services.关键词：leading time;consumers’ expected utility;two stages;logistics service;pricing strategy19|25|0更新时间：2026-09-15
- 摘要：While gradual typing improves program flexibility by combining static and dynamic typing, it also introduces runtime type inconsistencies and difficulties in attributing type errors. Although the existing Vigilance framework emphasizes the verification of type-history consistency, it cannot simultaneously support fine-grained error attribution. To address this limitation, this paper proposes a semantic extension of Vigilance based on a blame mechanism (hereinafter referred to as “Blame aware Vigilance semantics”). The extension introduces directional blame labels into the core intermediate language ICTL, enabling cast, monitor, and assert expressions to emit blame information that precisely identifies the source of a type mismatch as either the value (provider) or the context (consumer). Accordingly, the structures of the value log and type history are augmented to record blame information, and two core logical relations are extended to formally capture the triggering and propagation of blame. We prove that this blame aware Vigilance semantics not only preserves the original Vigilance guarantee, that every runtime value satisfies all its historical type constraints, but also provides precise and consistent blame attribution. An illustrative case study on a typical image-processing pipeline demonstrates that the blame aware Vigilance semantics can accurately pinpoint the responsible party (e.g., a faulty function implementation or an incorrect calling context) using blame labels and type history, significantly outperforming the traditional Vigilance approach, which merely reports type mismatch without attribution. This advancement substantially improves debugging efficiency and practical applicability.关键词：gradual typing;blame calculus;semantics;logical relations;type inference5|42|0更新时间：2026-09-15
- 摘要：Cultural landscapes in cold-region traditional villages are subject to the combined effects of severe cold climates, freeze-thaw processes, and socioeconomic transformation over the long term, resulting in pronounced contextual dependence, systemic interrelatedness, and structural heterogeneity in their vulnerability. In response to the insufficient attention to environmental adaptability in cold regions, inadequate identification of vulnerability structures, and weak linkage between assessment outcomes and differentiated governance in existing studies, this study takes 10 nationally designated traditional villages in Mudanjiang City as the research subjects. Environmental adaptability in cold regions is incorporated as the regional analytical context, and a cultural landscape vulnerability assessment system is established based on the Pressure-State-Response (PSR) model. The combination weighting method, composite index method, and factor contribution analysis are employed to identify the vulnerability level, structural characteristics, and dominant constraints of cultural landscapes, followed by typological classification of cultural landscape vulnerability and differentiated governance analysis. The results show that: 1) the vulnerability of cultural landscapes exhibits a structural pattern characterized by “pressure-driven, state-sensitive, and response-constrained”. Governmental and villager responses constitute the key constraints on overall cultural landscape vulnerability. The external disturbance induced by rural tourism development is greater than that associated with rapid urbanization, while the vulnerability of intangible cultural landscapes is generally higher than that of tangible cultural landscapes. 2) Cultural landscapes as a whole are characterized by a low level of vulnerability, although substantial inter-village variation exists, forming a spatial pattern dominated by low-vulnerability villages with localized occurrence of medium-vulnerability villages. Low-vulnerability villages are primarily distributed in and around Bohai Town, and their spatial differentiation is closely associated with locational conditions, resource endowments, development intensity, and the allocation of governance resources. For the three identified vulnerability types—Response-dominant (R), Pressure-State composite (PS), and State-Response composite (SR)-differentiated governance strategies are proposed, including strengthening grassroots conservation governance and optimizing resource allocation, controlling disturbance from rural tourism development and restoring degraded cultural landscapes, and enhancing landscape state restoration and coordination with villager participation, respectively. This study incorporates environmental adaptability in cold regions into the regional explanatory framework of cultural landscape vulnerability in traditional villages and establishes an analytical pathway of “cold-region environmental context–pressure effects-state changes-response regulation-vulnerability characterization.” The findings provide a theoretical basis for adaptive conservation and refined governance of cultural landscapes in cold-region traditional villages.关键词：cold-region traditional villages;cultural landscape vulnerability;PSR model;structural identification;typological classification;differentiated governance;Mudanjiang City23|26|0更新时间：2026-09-14
- 摘要：To address the unclear deformation mechanisms and insufficient displacement prediction accuracy of step-like landslides under the combined influence of rainfall and reservoir water level fluctuations, this study focuses on Baijiabao landslide on the right bank of the Xiangxi River in the Three Gorges Reservoir area. Through geological surveys, multi-source monitoring, and numerical simulations, the deformation evolution mechanisms of the landslide were systematically analyzed, and a displacement-oriented hybrid prediction model based on Variational Mode Decomposition (VMD), Improved Dung Beetle Optimization (IDBO), and Long Short-Term Memory (LSTM), was developed. The research results indicate that: 1) The mismatch between the reservoir water level decline rate and the seepage response rate of the low-permeability sliding mass leads to increased excess pore water pressure within the slip zone, which is the main cause of triggering step-like displacement; 2) The influence of rainfall on landslide deformation is relatively weak, mainly manifested as a superimposed effect on deformation at the rear edge of the landslide; 3) Reservoir water level fluctuations alter the seepage field in the slope, resulting in a staged variation characteristic of “enhancement-stabilization-weakening” in landslide stability; 4) The hybrid model (VMD-IDBO-LSTM) constructed based on displacement data of Baijiabao landslide demonstrates significantly better prediction accuracy than comparative models (VMD-LSTM and VMD-DBO-LSTM), with a root mean square error (RMSE) of 8.894 mm and a correlation coefficient (R²) of 0.983, effectively predicting step-like landslides in the reservoir area.关键词：step-like landslide;Baijiabao landslide;reservoir water level fluctuation;deformation mechanism;displacement prediction1084|100|0更新时间：2026-09-14
- 摘要：Split learning (SL) divides a machine learning model into several parts, allowing the client and server to collaboratively train and infer. However, studies showed that SL is vulnerable to data reconstruction attacks (DRAs). Differential privacy（DP） can mitigate the aforementioned threats, but current DP schemes for SL have two shortcomings: first, they do not provide formal DP guarantees for the entire SL training process; second, the differential privacy noise is too high on high-dimensional smashed data, leading to a significant drop in image classification accuracy. To protect the entire training process, we propose DPSplit, a DP training scheme for SL that adds DP noise to smashed data to protect the entire training process and employs an allocation algorithm to increase the per-iteration privacy budget under a total privacy budget, thereby reducing noise. To address the excessive noise problem, we propose R1SMG-E, a DP mechanism for high-dimensional matrices that uses rank-1 covariance matrix noise injection to make noise scale independent of dimension. As an independent differential privacy mechanism, R1SMG-E can be applied to DPSplit and other training schemes to further reduce noise. Experiments show that models trained under the DPSplit scheme with R1SMG-E maintain high utility and effectively resist various data reconstruction attacks.关键词：differential privacy（DP）;split learning（SL）;data reconstruction attacks24|67|0更新时间：2026-09-14
- 摘要：In this paper, the reaction mechanism between phenylalanine calcium (Phe→Ca2+) and hydroxyl radical (OH·) in free water and bound water environments at 310.15 K and 1.013×105 Pa in vivo was investigated using density functional theory (DFT). The reaction proceeds via three channels: H-abstraction by OH·, OH· addition, and single-electron transfer. Calculations indicate that in the free water environment, the favorable H-abstraction reactions are significantly exothermic processes with energy barriers ranging from 35.2 to 49.6 kJ/mol; the favorable addition reactions are exothermic processes with energy barriers ranging from 27.7 to 45.4 kJ/mol. In the bound water environment, the favorable H-abstraction reactions are exothermic processes with energy barriers ranging from 40.0 to 45.8 kJ/mol; the favorable addition reactions are exothermic processes with energy barriers ranging from 38.8 to 45.5 kJ/mol. In both free water and bound water environments, the electron transfer process is thermodynamically and kinetically unfavorable. The results indicate that Phe→Ca2+ can scavenge OH· radicals via both H-abstraction and addition pathways in free water and bound water environments in vivo.关键词：phenylalanine calcium;hydroxyl radical;density functional theory;free water;bound water;free energy barrier;amino acids12|35|0更新时间：2026-09-14
- 摘要：Certificate Transparency audits digital certificates through a public logging mechanism, where auditing the signed certificate timestamp effectively prevents certificate misissuance, which serves as a crucial mechanism for ensuring network security. However, existing schemes for protecting query privacy face inherent limitations: anonymity-based approaches rely on additional trust assumptions and fail to provide strong privacy guarantees, while cryptography-based approaches suffer from excessive server storage overhead or inefficient dynamic maintenance in large-scale scenarios. To address this, this paper proposes a signed certificate timestamp auditing (SCTA) system based on private information retrieval. By designing a novel primitive called an extended binary fuse filter, the system combines the efficient query capabilities of binary fuse filters with the dynamic deletability of counting Bloom Filters. This enables rapid verification of static data and real-time updates for dynamic data. By integrating private information retrieval with deletion log management, it supports efficient data insertion, deletion, and persistent deletion while protecting client query privacy. Formal analysis proves that SCTA satisfies client query privacy under the semi-honest adversary model. Experimental results demonstrate that SCTA system significantly reduces server storage overhead while maintaining controllable computational overhead, offering comprehensive advantages in both performance and security compared to the baseline approach.关键词：certificate transparency;signed certificate timestamp;private information retrieval;binary fuse filter15|61|0更新时间：2026-09-07
- 摘要：To investigate the contamination characteristics of the herbicide glyphosate (PMG) in a typical tributary bay of the Three Gorges Reservoir Area, a four-season sampling survey was conducted at 9 sites in the Xiangxi Bay from March to December 2025. The spatiotemporal distribution patterns, environmental responses were analysed and ecological risk of PMG and AMPA were further evaluated using the risk quotient (RQ) approach. The results showed that PMG was detected at all sampling sites throughout the four seasons with concentrations ranging from 0.01 to 4.54 μg/L. The detection rate of AMPA was 100% in spring, summer, and autumn, but decreased to 66.67% in winter, with concentrations ranging from 0 to 2.78 μg/L. PMG and AMPA exhibited distinct spatiotemporal decoupling patterns. PMG concentrations followed a decreasing seasonal trend from spring to winter and displayed a spatial distribution pattern with higher levels in upstream regions and lower levels in downstream and confluence areas. In contrast, AMPA concentrations peaked in summer and exhibited a reverse spatial accumulation pattern, with higher concentrations in the confluence and downstream areas than in upstream regions. Environmental response analysis revealed that the occurrence of PMG was mainly influenced by agricultural non-point source inputs, rainfall-driven runoff transport, and physicochemical conditions of the water environment, while the generation and accumulation of AMPA were highly regulated by microbial transformation processes and hydrodynamic conditions within the reservoir. Ecological risk assessment indicated that PMG reached a medium risk level (0.1≤RQ<1) at some sites during spring and summer and was the major contributor to the glyphosate-related ecological riskin the study area, while AMPA posed negligible ecological risk (RQ<0.01). These findings elucidate the migration and transformation patterns of PMG and AMPA in typical tributaries of the Three Gorges Reservoir Area and provide a scientific basis for the precise management of non-point source pollution and the prevention of water environmental risks in the reservoir area.关键词：Three Gorges Reservoir;Xiangxi Bay;glyphosate (PMG);AMPA;spatiotemporal distribution;ecological risk17|47|0更新时间：2026-08-17
- 摘要：To meet the demand for high conductivity and mechanical toughness of hydrogels in flexible electronics, the rapid preparation of tough polyacrylic acid (PAA) hydrogels in a one-step process was achieved by an efficient redox initiation system using ascorbic acid (Laa) and ammonium persulfate (APS) and the synergistic effect of the coordination effect of Zr4+ and the Hofmeister effect of (NH4)2SO4. The gelation time was only 100 s. The effects of Zr4+ amount and (NH4)2SO4 concentration on the hydrogel’s properties were systematically investigated. The PAA hydrogels prepared under optimized conditions exhibit an excellent comprehensive performance: they have outstanding mechanical properties with a stress of up to 2.72 MPa and a strain of up to 1 429.7%; meanwhile, they possess an ionic conductivity of 73.46 mS/cm; in addition, they also exhibit excellent fatigue resistance and shape recovery ability. Due to these properties, the conductive hydrogel holds great promise for applications in flexible energy storage devices, flexible wearable sensors, soft robotics actuation, and biomedical monitoring.关键词：hydrogels;flexible electronic devices;mechanical properties;conductivity;redox initiating system;coordination interaction;Hofmeister effect16|57|0更新时间：2026-08-10
- 摘要：Ion concentration variations are closely related to physiological homeostasis and disease onset and progression. Accurate and continuous ion monitoring provides essential information for health management and early disease screening. Flexible transistors have emerged as promising platforms for next-generation ion sensors owing to their excellent electrical performance and mechanical compliance. This review focuses on two major device types: field-effect transistors (FETs) and organic electrochemical transistors (OECTs), and systematically summarizes recent advances in flexible ion sensing. We compare inorganic and organic FETs and OECTs in terms of material systems, device architectures, sensing performance, and summarize their applications in wearable electronics and bioelectronics. Ultimately, we discuss current challenges and outline future development directions.关键词：field-effect transistors;flexibility;ion sensing;electrical performance;mechanical compliance52|64|0更新时间：2026-07-29
- 摘要：Red teaming is essential for ensuring the safe deployment of Large Language Models (LLMs) in practical applications. However, existing automated red teaming methods often overlook the generalization ability of attack samples and typically rely on a single jailbreak strategy, with limited exploration of mechanisms for integrating and coordinating multiple strategies, making them insufficient for practical needs. To address these challenges, this paper proposes an automated red teaming framework for LLMs that integrates multiple jailbreak strategies. The framework leverages an attack model to generate semantically diverse attack samples and incorporates 22 jailbreak strategies in a modular design, supporting flexible composition and dynamic extensibility, thereby enabling comprehensive red teaming against target models. In addition, the framework utilizes the red teaming data to further enhance the security alignment of the target model. Experiments show that the generated attack samples achieve strong performance in both violation category coverage and textual similarity, with an attack success rate (ASR) of 94.29%, surpassing the state-of-the-art method by 17.37 percentage points. Furthermore, after alignment optimization, the ASR is significantly reduced to 14.00%, validating the effectiveness and practical value of the proposed framework in both red teaming and model alignment.关键词：large language model;red teaming;jailbreak strategy;model alignment14|67|0更新时间：2026-07-28
- 摘要：Dialect speech exhibits significant variations in pronunciation patterns, acoustic distributions, and linguistic structures, which poses challenges to end-to-end speech recognition models, especially under low-resource conditions. This paper proposes a Sichuan dialect speech recognition method based on multi-scale convolutional Conformer by optimizing both encoding and decoding stages. The model is composed of a multi-scale convolutional encoder and a hybrid decoder jointly modeled by connectionist temporal classification (CTC) and attention mechanism. In the encoding stage, a multi-scale convolutional structure is incorporated into the convolution module of Conformer to achieve parallel extraction of local acoustic features across different temporal scales, and a gated fusion mechanism is employed to adaptively integrate multi-scale representations, enhancing the modeling of complex dialect speech characteristics. In the decoding stage, a hybrid decoding architecture of CTC and bidirectional Transformer is adopted, integrating a specific N-gram language model trained on Sichuan dialect text, and a dynamic attention rescoring strategy is proposed to adaptively adjust the fusion weights between acoustic and language model scores based on decoding path statistics. Experimental results on a real Sichuan dialect speech dataset show that the proposed method reduces the character error rate by approximately 17%, which demonstrates the effectiveness of the proposed approach for Sichuan dialect speech recognition.关键词：speech recognition;dialect speech recognition;Conformer;end-to-end model38|110|0更新时间：2026-07-23
- 摘要：The incorporation of Knowledge Graphs (KGs) has effectively alleviated the data sparsity and cold-start issues in recommender systems. However, collaborative KGs in real-world scenarios inevitably suffer from long-tail distributions and interaction noise. Conventional graph contrastive learning augmentation strategies often rely on heuristic random topological deletions, which easily disrupt the original collaborative network structure. To address this, we propose an SVD（Singular Value Decomposition）-based View and Residual Contrastive Learning framework for knowledge-graph recommendation, referred to as SVRCL. The model performs SVD on the entire representation matrix in the spectral space. By injecting adaptive Gaussian noise directly into the singular value distribution and reconstructing a global contrastive view, it suppresses high-frequency random noise while preserving core topological connectivity. Furthermore, to tackle the sparse gradient signal problem in traditional ranking optimization, we design a residual-reweighted BPR loss based on Sigmoid hardness. This strategy dynamically quantifies the hardness of negative samples to accurately allocate gradient weights, thereby optimizing the discriminative boundary of the recommender system. Extensive experiments on two publicly available datasets, Last-FM and Book-Crossing, show that SVRCL outperforms existing mainstream baseline models across multiple core performance metrics. On the highly sparse Book-Crossing dataset, compared to the runner-up model LightKG, SVRCL improves Recall@10, NDCG@10, and MRR@10 by 12.39%, 7.26%, and 5.71%, respectively, with lower time and memory overhead, achieving an effective balance between recommendation robustness and computational efficiency in sparse scenarios.关键词：graph neural network;knowledge graph;recommendation algorithm;contrastive learning;singular value decomposition10|75|0更新时间：2026-07-21
- Tel：027-68756952 Email：whdz@whu.edu.cn
- Technical support is provided by Beijing Founder electronics co., LTD 京ICP备09064830号-19
京公网安备11010802024621 - It is recommended to read the content of this site in Chrome&IE9+. Please switch to extreme mode in browser 360.
- Cookies We use cookies to help provide and enhance our service and tailor content. By continuing, you agree to the use of cookies.
- Total Visits：157145 Visits Today：109 Online Users：9
Bulk Citation
0