Loading...

Table of Content

    11 November 2026, Volume 12 Issue 11
    A Framework for Agent Security Governance
    Gu Hongxun
    2026, 12(11):  978-983.  DOI: 10.12379/j.issn.2096-1057.2026.11.01
    Asbtract ( )   PDF (574KB) ( )  
    References | Related Articles | Metrics
    Agents are moving from proof of concept to large-scale deployment. However, their autonomous execution can easily lead to behavioral runaway, algorithmic black boxes make decision processes difficult to trace, and cross-domain collaboration blurs accountability boundaries. Accordingly, this paper constructs a “five objectives–four phases–nine measures” security governance framework, which takes trusted identity, controllable authorization, auditable process, traceable accountability and manageable scenario as the objectives, risk observation, runtime management, behavior intervention and capability improvement as the closed-loop phases, and nine capabilities including interaction detection and context protection as the measures. An L1~L5 graded behavior intervention mechanism is further proposed, providing a reusable governance paradigm for full lifecycle security management of enterprise agents.
    A Multimodal Feature Fusion Approach for Vulnerability Detection Based on Large Models
    Zhao Xinjian, Tao Chiyu, Chen Shi, Kuang Boyu, Wu Zicheng, Fu Anmin
    2026, 12(11):  984-992.  DOI: 10.12379/j.issn.2096-1057.2026.11.02
    Asbtract ( )   PDF (949KB) ( )  
    References | Related Articles | Metrics
    With the widespread adoption of open-source software, the number of vulnerabilities has increased sharply. The traditional detection methods suffer from high false positive rates and difficulty in capturing complex dependencies. To address these issues, this paper proposes a large model-based multi-modal feature fusion vulnerability detection method named MMFVD. The method first constructs a Multi-functional Code Property Graph (M-CPG), which retains key structural information through program slicing and pruning. Next, pre-trained sequence models and graph models are used to extract semantic features and structural features of the code respectively, with online knowledge distillation applied to optimize feature representation. Finally, a dynamic attention mechanism is adopted to achieve adaptive fusion of the features, and a classifier is employed to complete vulnerability prediction. On four public datasets, MMFVD is compared with 10 baseline methods, which verifies the significant advantages of MMFVD in detection accuracy, robustness, and generalization.
    Smart Contract Deep Learning Detection Model for Feature Fusion
    Gong Yixuan, Wu Kaimin, Ma Hailiang
    2026, 12(11):  993-1000.  DOI: 10.12379/j.issn.2096-1057.2026.11.03
    Asbtract ( )   PDF (1143KB) ( )  
    References | Related Articles | Metrics
    With the rapid development of blockchain technology, smart contract vulnerability detection has become a critical task for ensuring the security of blockchain systems. Existing advanced methods typically transform bytecode into graph structures and apply graph neural networks for feature learning, but they insufficiently exploit the semantic and structural characteristics of source code. To address this issue, this paper proposes ACF-VDF, a vulnerability detection framework that fuses Abstract Syntax Tree (AST) semantic features with Control Flow Graph (CFG) structural features. Specifically, the method constructs ASTs from smart contract source code to obtain syntax feature vectors that capture hierarchical and semantic information; meanwhile, contract bytecode is transformed into CFGs, and through global pooling and node feature extraction, structural feature vectors that reflect execution paths and structural dependencies are generated. The two feature vectors are jointly modeled and fed into a Transformer-based deep neural network classifier to detect vulnerabilities in smart contracts. Experimental results demonstrate that the proposed method improves the F1 score by 2.10%–34.28% compared with seven state-of-the-art detection methods, significantly boosting both accuracy and generalization performance in vulnerability detection.
    Lightweight AI-Generated Text Detection Model Fusing Multi-dimensional Features and Contrastive Learning
    Zhi Luping, Fang Liying, Cai Miao
    2026, 12(11):  1001-1010.  DOI: 10.12379/j.issn.2096-1057.2026.11.04
    Asbtract ( )   PDF (834KB) ( )  
    References | Related Articles | Metrics
    With the rapid development of generative large language models, the texts generated by these models are increasingly similar to human texts. Efficient detection is required to avoid misleading the public. Nevertheless, sophisticated and accurate detection methods suffer from high computational resource consumption and long inference time. To meet this requirement, this paper designs a lightweight detection scheme. Specifically, Sentence-BERT is adopted to capture implicit semantic logical correlations of texts; the Natural Language Toolkit (NLTK) library is used to extract sentiment and statistical features for constructing a shallow feature set; an attention-weighted mechanism is introduced for adaptive fusion of the two categories of shallow features to strengthen the representation capability of critical shallow information; after dimension-wise concatenation of fused shallow features and deep semantic representations, contrastive learning is utilized to optimize joint features, and Support Vector Machine (SVM) is applied for final classification. The model adopts a coupled architecture of deep semantic representation and lightweight classification, which fundamentally reduces computational resource consumption and shortens inference time. It achieves detection accuracies of 92% and 90% on Chinese and English test sets, realizing the balance between detection accuracy and efficiency.
    Research on Collaborative Governance Mechanisms for the Dissemination of AI Deepfake Videos
    Ma Xiaofei, Wang Jia, Wang Zhongyu
    2026, 12(11):  1011-1019.  DOI: 10.12379/j.issn.2096-1057.2026.11.05
    Asbtract ( )   PDF (751KB) ( )  
    References | Related Articles | Metrics
    The rapid advancement of generative AI technology has given rise to deepfake videos, which pose a series of security risks. Effectively regulating the proliferation of deepfake videos has become a critical issue concerning information security. This paper is based on information ecosystem theory and constructs a four-party game model involving the government, video platforms, video disseminators, and video recipients, followed by numerical simulation analysis. The study found that: the government and video platforms should break through linear regulatory thinking to promote collaborative governance, rather than relying on a single external force for governance; Increasing the severity of rewards and penalties for governments and video platforms can effectively promote the governance of deepfake video dissemination, with video platforms offering rewards that have an effective threshold; Reducing the excess profit margins of video platforms, video disseminators, and video recipients can promote collaborative governance among the four parties, with government behavior remaining largely unaffected. This indicates that the government should fully leverage its role as a facilitator to promote the formation of a more rational market mechanism.
    Security Risks and Governance Paths for AI-Agent Outputs from a Copyright Perspective
    Ren ZuLiang
    2026, 12(11):  1020-1030.  DOI: 10.12379/j.issn.2096-1057.2026.11.06
    Asbtract ( )   PDF (626KB) ( )  
    References | Related Articles | Metrics
    AI agents have transformed the logic of data training, tool invocation, and content dissemination in generative artificial intelligence, creating behaviour-chain risks involving multiple actors. These risks are manifested in the extension of copyright infringement to material-invocation scenarios, the involvement of multiple participants in infringement risks, and the ecological damage caused by automated output of infringing content. The governance framework for generative AI content has traditionally relied on content moderation and platform model regulation, making it incapable of managing the risk or failing to interrupt the infringement behaviour chain. That results in three structural mismatches: between centralised control of knowledge and decentralised multi-actor risks, between content review and segmented behavioural control, and between ex-post notice-and-takedown mechanisms and the attribution of chain-based infringements. Accordingly, the regulatory focus should shift towards behavior chain control, establishing a responsibility allocation model based on behavioural chains to facilitate risk prevention under a multi-actor duty-of-care framework. Specifically, this entails users’ obligations to define tasks and confirm authorization, AI agent service providers’ end-to-end preventive obligations, third-party technology providers’ obligations to respect the boundaries of technical capabilities, and a responsibility attribution mechanism that runs through the entire behaviour chain.
    An Unsupervised Traffic Anomaly Detection Method for Unknown Threats
    Zhang Zhen, Sun ShuoChao, Liu XinPeng
    2026, 12(11):  1031-1041.  DOI: 10.12379/j.issn.2096-1057.2026.11.07
    Asbtract ( )   PDF (965KB) ( )  
    References | Related Articles | Metrics
    With the continuous growth of network services and the widespread adoption of encrypted communications, complex protocols, and dynamic interaction patterns, network traffic anomaly detection is facing increasing challenges, including the emergence of unknown threats, the scarcity of attack labels, and time-varying data distributions. To address these issues, this paper proposes an unsupervised network traffic anomaly detection method for unknown threats, named Multi-view Unsupervised Self-supervised Traffic Anomaly Detection (MUST-AD). Taking network flows as detection objects, the proposed method constructs multi-view representations by jointly modeling statistical behavior, temporal evolution, and communication relationships. It then establishes a latent representation space for normal traffic through a self-supervised learning strategy that combines masked reconstruction and cross-view contrastive learning. In addition, normal prototype organization, joint anomaly scoring, and drift-aware online updating are integrated to achieve comprehensive detection of unknown anomalous traffic. Experiments are conducted on the UNSW-NB15, CIC-IDS2017, and MAWIFlow datasets. The results show that, in-domain static detection scenarios, MUST-AD improves Accuracy and AUC by 0.1%~0.3% and improves F1 by approximately 0.6~1.0 percentage points compared with baseline methods. Under cross-dataset settings, the AUC is improved by 2~3 percentage points. On MAWIFlow, the AUC reaches 0.915~0.920. In cross-time generalization scenarios, the AUC reaches 0.84, outperforming the compared methods with AUC values of 0.79 and 0.61. The proposed method can simultaneously maintain detection performance, cross-scenario generalization capability.
    IoT Intrusion Detection Based on Class-Aware and Weighted Ensemble
    Chen Liangchen, Fu Deyin, Zhang Shuo, He Boshu, Lu Zhigang, Liu Baoxu
    2026, 12(11):  1042-1051.  DOI: 10.12379/j.issn.2096-1057.2026.11.08
    Asbtract ( )   PDF (819KB) ( )  
    References | Related Articles | Metrics
    To address high-dimensional feature redundancy, class imbalance, and limited detection performance in complex attack IoT (Internet of Things)scenarios, this study proposes an IoT Intrusion Detection method integrating Class-aware Feature selection and Weighted Ensemble learning(IID-CFWE). First, a Class-aware Adaptive recursive Feature Selection(CA-AFS) mechanism is developed by jointly considering global, class-aware, and boundary-sample feature importance, thereby reducing redundant features and enhancing the discrimination of minority attack classes. Second, a heterogeneous ensemble learning model combining random-search optimization with weighted soft voting is constructed. Random search is employed to optimize the hyperparameters of base classifiers, while the class probability outputs of base classifiers are weighted according to their validation performance and fused by weighted soft voting, enabling the model to exploit their complementary discriminative capabilities. The class-probability outputs of the base classifiers are then aggregated through weighted soft voting to improve detection performance and classification stability. Extensive experiments are conducted on the public N-BaIoT and CIC-IDS-2017 datasets. The results indicate that the proposed method effectively improves intrusion detection performance while reducing feature redundancy, providing an efficient solution for IoT network security.
    Design of a Zero Trust-Based Security Architecture for Large-Scale Mobile Government Applications
    Liu Bei, Wang Xudong, Zhao Binqi
    2026, 12(11):  1052-1060.  DOI: 10.12379/j.issn.2096-1057.2026.11.09
    Asbtract ( )   PDF (715KB) ( )  
    References | Related Articles | Metrics
    With the widespread adoption of mobile internet and the rapid advancement of mobile intelligent terminal technologies, promoting more government services to be accessible via mobile and handheld devices has become a critical approach to enhancing government administrative efficiency and service effectiveness. Mobile government systems have also emerged as a vital component of digital government and digital institution construction. To ensure the security of mobile government systems and address the security risks arising from open interconnectivity, this paper proposes a zero-trust security architecture tailored for large-scale mobile government application scenarios. This architecture achieves comprehensive security protection across mobile terminals, mobile communications, mobile access, and mobile services, and employs a "distributed cluster deployment + intelligent dynamic routing" strategy to realize a dual high-availability architecture. The effectiveness and feasibility of the proposed architecture have been validated through construction practices and testing results on the mobile side of a provincial integrated government office platform.
    Research on Maturity Assessment Model of Trusted Data Circulation Technology
    Wang Xiaodong, Dong Chao
    2026, 12(11):  1061-1068.  DOI: 10.12379/j.issn.2096-1057.2026.11.10
    Asbtract ( )   PDF (639KB) ( )  
    References | Related Articles | Metrics
    Trusted data circulation is an important prerequisite for building a unified data market and holds significant importance for ensuring data supply availability, enabling smooth circulation, facilitating effective utilization, and guaranteeing data security. Currently, data circulation faces a complex landscape shaped by legal compliance requirements, the awakening of data awareness among market entities, innovation in digital intelligence technologies, and the superposition of multiple attributes of data. Traditional circulation models are inadequate to adapt to future large-scale data circulation scenarios. New technologies and solutions for trusted data circulation are continuously emerging. How to scientifically evaluate these new technologies and solutions has become an urgent issue. This paper systematically sorted out the practical dilemmas and development trends of trusted data circulation. It constructed a six-dimensional "ACCESS" evaluation model, with core dimensions including process auditability, circulation compliance, technical compatibility, trusted circulation efficiency, data security, and technical standardization. This model provides an analytical framework and practical guidance for promoting technological innovation and maturity assessment of trusted data circulation.
    The Impact of Description-Behavior Inconsistency on Tool-Calling Safety in LLM Agents
    Qian Hanwei, Ma Zhuo, Xia Lingling, Guan Junqi, Wang Qun
    2026, 12(11):  1069-1077.  DOI: 10.12379/j.issn.2096-1057.2026.11.11
    Asbtract ( )   PDF (704KB) ( )  
    References | Related Articles | Metrics
    LLM agents primarily rely on developer-provided natural-language descriptions when selecting tools, yet current agent frameworks lack a mechanism that guarantees consistency between tool descriptions and actual behavior. To study the safety impact of description distortion, we construct a controlled benchmark covering 20 danger categories with multiple distortion levels and channel-ablation conditions. Across three model families, greater description distortion is significantly associated with a higher Unsafe Invocation Rate (UIR). Under false reassurance, UIR reaches 2.4 to 5.4 times the honest-description baseline, and over 80% of unsafe calls contain no explicit risk-related wording in their rationales. Further analysis shows that danger signals in tool names and parameter schemas help reduce UIR, while two lightweight description-based mitigations reduce GPT's UIR under false reassurance by 37% to 47%. These results indicate that danger information should be expressed explicitly across the name, description and parameter-schema channels in tool design.