Journal of Information Security and Applications最新文献_第6页

Enhancing black-box membership inference attacks in federated learning 增强联邦学习中的黑盒成员推理攻击

IF 3.7 2区计算机科学 Q2 COMPUTER SCIENCE, INFORMATION SYSTEMS

Journal of Information Security and Applications

Pub Date : 2025-11-15 DOI: 10.1016/j.jisa.2025.104302

Qiang Shi, Luzhen Ren, Xinfeng He

With the widespread deployment of machine learning models in privacy-sensitive domains such as healthcare and finance, the risk of training data leakage has attracted increasing attention. As a fundamental approach for evaluating model privacy leakage, Membership Inference Attack (MIA) has been extensively studied in distributed learning scenarios such as Federated Learning (FL). However, under black-box settings, attackers face severe challenges, including the unavailability of real non-member samples and the inaccessibility of target model architectures, which limit the generalization and accuracy of existing methods. To address these limitations, this paper proposes a DCGAN-enhanced black-box MIA framework, whose innovations are reflected in three major aspects: (1) a discriminator-guided pseudo-sample filtering mechanism that ensures the authenticity and diversity of non-member data; (2) a multi-shadow-model softmax high-dimensional concatenation strategy, which fuses the softmax probability outputs from multiple shadow models to construct discriminative high-dimensional attack representations; and (3) a SMOTE-based balancing module designed to mitigate class imbalance and further improve the generalization of the attack model. The proposed framework significantly enhances the discriminative capability and robustness of black-box MIAs without accessing the internal parameters or training procedures of the target model. Extensive experiments demonstrate that our method consistently outperforms state-of-the-art baselines across multiple federated learning protocols (FedAvg, FedMD, and FedProx) and benchmark datasets (CIFAR-10, CIFAR-100, Fashion-MNIST, and SVHN), achieving an accuracy of 0.9897, an AUC of 0.9899, and a TPR@FPR=1% of 0.9967. These results verify the robustness, generalizability, and wide applicability of the proposed framework, providing a systematic and scalable solution for privacy evaluation in federated learning environments.

随着机器学习模型在医疗保健和金融等隐私敏感领域的广泛应用，培训数据泄露的风险越来越受到关注。隶属关系推理攻击（MIA）作为评估模型隐私泄露的基本方法，在联邦学习（FL）等分布式学习场景中得到了广泛的研究。然而，在黑盒设置下，攻击者面临着严峻的挑战，包括真实非成员样本的不可获得性和目标模型体系结构的不可访问性，这限制了现有方法的泛化和准确性。针对这些局限性，本文提出了一种基于dcgan的黑箱MIA框架，其创新主要体现在三个方面：(1)采用了鉴别器引导的伪样本过滤机制，保证了非成员数据的真实性和多样性；(2)多阴影模型softmax高维拼接策略，融合多个阴影模型的softmax概率输出，构建判别性高维攻击表征；(3)基于smote的平衡模块，旨在缓解类不平衡，进一步提高攻击模型的泛化性。该框架在不访问目标模型内部参数或训练过程的情况下，显著提高了黑箱MIAs的判别能力和鲁棒性。广泛的实验表明，我们的方法在多个联邦学习协议（fedag、FedMD和FedProx）和基准数据集（CIFAR-10、CIFAR-100、fashionon - mnist和SVHN）上始终优于最先进的基线，实现了0.9897的准确率，0.9899的AUC和TPR@FPR=1%的0.9967。这些结果验证了所提出框架的鲁棒性、泛化性和广泛适用性，为联邦学习环境中的隐私评估提供了系统和可扩展的解决方案。

{"title":"Enhancing black-box membership inference attacks in federated learning","authors":"Qiang Shi, Luzhen Ren, Xinfeng He","doi":"10.1016/j.jisa.2025.104302","DOIUrl":"10.1016/j.jisa.2025.104302","url":null,"abstract":"<div><div>With the widespread deployment of machine learning models in privacy-sensitive domains such as healthcare and finance, the risk of training data leakage has attracted increasing attention. As a fundamental approach for evaluating model privacy leakage, Membership Inference Attack (MIA) has been extensively studied in distributed learning scenarios such as Federated Learning (FL). However, under black-box settings, attackers face severe challenges, including the unavailability of real non-member samples and the inaccessibility of target model architectures, which limit the generalization and accuracy of existing methods. To address these limitations, this paper proposes a DCGAN-enhanced black-box MIA framework, whose innovations are reflected in three major aspects: (1) a discriminator-guided pseudo-sample filtering mechanism that ensures the authenticity and diversity of non-member data; (2) a multi-shadow-model softmax high-dimensional concatenation strategy, which fuses the softmax probability outputs from multiple shadow models to construct discriminative high-dimensional attack representations; and (3) a SMOTE-based balancing module designed to mitigate class imbalance and further improve the generalization of the attack model. The proposed framework significantly enhances the discriminative capability and robustness of black-box MIAs without accessing the internal parameters or training procedures of the target model. Extensive experiments demonstrate that our method consistently outperforms state-of-the-art baselines across multiple federated learning protocols (FedAvg, FedMD, and FedProx) and benchmark datasets (CIFAR-10, CIFAR-100, Fashion-MNIST, and SVHN), achieving an accuracy of 0.9897, an AUC of 0.9899, and a TPR@FPR=1% of 0.9967. These results verify the robustness, generalizability, and wide applicability of the proposed framework, providing a systematic and scalable solution for privacy evaluation in federated learning environments.</div></div>","PeriodicalId":48638,"journal":{"name":"Journal of Information Security and Applications","volume":"96 ","pages":"Article 104302"},"PeriodicalIF":3.7,"publicationDate":"2025-11-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145520857","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Cybersecurity Digital Twins: Concept, blueprint, and challenges for multi-ownership digital service chains 网络安全数字孪生：多所有权数字服务链的概念、蓝图和挑战

IF 3.7 2区计算机科学 Q2 COMPUTER SCIENCE, INFORMATION SYSTEMS

Journal of Information Security and Applications

Pub Date : 2025-11-15 DOI: 10.1016/j.jisa.2025.104299

M. Repetto

The growing level of interconnectedness of digital services and infrastructures creates tight and recursive security inter-dependencies between their providers. However, cybersecurity operations remain highly fragmented, since common tasks like disclosing vulnerabilities, reporting alerts, and suggesting remediation are largely restricted within the boundaries of the administrative domain of each provider, while cooperation is usually limited to paperwork and human interactions. This practice has already demonstrated to be inadequate and risky, because it cannot effectively address multi-step attacks and kill chains that propagate across multiple domains.

In this position paper, we elaborate on the concept, blueprint, and usage of a Cyber-security Digital Twin that models and captures the security posture of such interconnected systems. Differently from existing models, our work explicitly addresses the challenges brought by multi-ownership, by focusing on the overall architecture to build cooperative, agile, adaptive and autonomous processes for threat hunting, detection of lateral movements, and eradication of attacks among multiple domains. For this reason, our framework takes into account the necessary federation mechanisms that address trust and confidentiality concerns.

数字服务和基础设施的互联程度不断提高，在它们的提供商之间产生了紧密的、递归的安全相互依赖关系。然而，网络安全运营仍然高度分散，因为披露漏洞、报告警报和建议补救等常见任务在很大程度上限制在每个提供商的管理领域范围内，而合作通常仅限于文书工作和人际互动。这种做法已经被证明是不充分和有风险的，因为它不能有效地处理跨多个域传播的多步骤攻击和杀伤链。在这份意见书中，我们详细阐述了网络安全数字孪生模型的概念、蓝图和用法，该模型可以模拟和捕获此类互联系统的安全状态。与现有模型不同，我们的工作明确解决了多所有权带来的挑战，通过关注整体架构来构建合作，敏捷，自适应和自主的过程，用于威胁狩猎，检测横向移动，并消除多个领域的攻击。出于这个原因，我们的框架考虑了解决信任和机密性问题的必要联合机制。

{"title":"Cybersecurity Digital Twins: Concept, blueprint, and challenges for multi-ownership digital service chains","authors":"M. Repetto","doi":"10.1016/j.jisa.2025.104299","DOIUrl":"10.1016/j.jisa.2025.104299","url":null,"abstract":"<div><div>The growing level of interconnectedness of digital services and infrastructures creates tight and recursive security inter-dependencies between their providers. However, cybersecurity operations remain highly fragmented, since common tasks like disclosing vulnerabilities, reporting alerts, and suggesting remediation are largely restricted within the boundaries of the administrative domain of each provider, while cooperation is usually limited to paperwork and human interactions. This practice has already demonstrated to be inadequate and risky, because it cannot effectively address multi-step attacks and kill chains that propagate across multiple domains.</div><div>In this position paper, we elaborate on the concept, blueprint, and usage of a Cyber-security Digital Twin that models and captures the security posture of such interconnected systems. Differently from existing models, our work explicitly addresses the challenges brought by multi-ownership, by focusing on the overall architecture to build cooperative, agile, adaptive and autonomous processes for threat hunting, detection of lateral movements, and eradication of attacks among multiple domains. For this reason, our framework takes into account the necessary federation mechanisms that address trust and confidentiality concerns.</div></div>","PeriodicalId":48638,"journal":{"name":"Journal of Information Security and Applications","volume":"96 ","pages":"Article 104299"},"PeriodicalIF":3.7,"publicationDate":"2025-11-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145520858","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Blockchain-based threshold proxy re-encryption scheme with zero-knowledge proofs for confidential and verifiable IoT networks 基于区块链的阈值代理再加密方案，具有零知识证明，用于机密和可验证的物联网网络

IF 3.7 2区计算机科学 Q2 COMPUTER SCIENCE, INFORMATION SYSTEMS

Journal of Information Security and Applications

Pub Date : 2025-11-13 DOI: 10.1016/j.jisa.2025.104300

Vinay Rishiwal , Ved Prakash Mishra , A. Jayanthiladevi , Vinay Maurya , Udit Agarwal , Mano Yadav

The rapid proliferation of interconnected devices within the Internet of Things (IoT) continues to generate vast amounts of sensitive, context-rich data, raising significant concerns regarding data confidentiality, verifiability, trust management, and systemic resilience. Traditional IoT network architectures typically rely on centralised third-party entities. This reliance creates single points of failure and elevates the risk of unauthorised data access. To address these limitations, this paper proposes a confidential and verifiable IoT network based on a decentralised security architecture that integrates blockchain with proxy re-encryption. The framework uses threshold cryptography and zero-knowledge proofs to enable privacy-preserving transformations of ciphertext across consensus nodes. This design protects sensitive data while preserving transaction verifiability and integrity. As a result, the system effectively counters threats such as node collusion, Sybil attacks, and metadata leakage. Comprehensive simulations and performance evaluations underscore that the presented model substantially diminishes dependence on centralised proxies while delivering enhanced scalability, robust security, and increased trustworthiness, making it particularly well-suited for practical implementation in confidential IoT environments.

物联网（IoT）中互连设备的快速扩散继续产生大量敏感的、上下文丰富的数据，引起了对数据机密性、可验证性、信任管理和系统弹性的重大关注。传统的物联网网络架构通常依赖于集中式第三方实体。这种依赖造成了单点故障，并增加了未经授权访问数据的风险。为了解决这些限制，本文提出了一个基于分散安全架构的机密和可验证的物联网网络，该架构将区块链与代理重新加密集成在一起。该框架使用阈值密码学和零知识证明来实现跨共识节点的密文隐私保护转换。这种设计保护敏感数据，同时保持事务的可验证性和完整性。有效应对节点合谋、Sybil攻击、元数据泄露等威胁。综合模拟和性能评估强调，所提出的模型大大减少了对集中式代理的依赖，同时提供了增强的可扩展性、强大的安全性和更高的可信度，使其特别适合在机密物联网环境中实际实施。

{"title":"Blockchain-based threshold proxy re-encryption scheme with zero-knowledge proofs for confidential and verifiable IoT networks","authors":"Vinay Rishiwal , Ved Prakash Mishra , A. Jayanthiladevi , Vinay Maurya , Udit Agarwal , Mano Yadav","doi":"10.1016/j.jisa.2025.104300","DOIUrl":"10.1016/j.jisa.2025.104300","url":null,"abstract":"<div><div>The rapid proliferation of interconnected devices within the Internet of Things (IoT) continues to generate vast amounts of sensitive, context-rich data, raising significant concerns regarding data confidentiality, verifiability, trust management, and systemic resilience. Traditional IoT network architectures typically rely on centralised third-party entities. This reliance creates single points of failure and elevates the risk of unauthorised data access. To address these limitations, this paper proposes a confidential and verifiable IoT network based on a decentralised security architecture that integrates blockchain with proxy re-encryption. The framework uses threshold cryptography and zero-knowledge proofs to enable privacy-preserving transformations of ciphertext across consensus nodes. This design protects sensitive data while preserving transaction verifiability and integrity. As a result, the system effectively counters threats such as node collusion, Sybil attacks, and metadata leakage. Comprehensive simulations and performance evaluations underscore that the presented model substantially diminishes dependence on centralised proxies while delivering enhanced scalability, robust security, and increased trustworthiness, making it particularly well-suited for practical implementation in confidential IoT environments.</div></div>","PeriodicalId":48638,"journal":{"name":"Journal of Information Security and Applications","volume":"96 ","pages":"Article 104300"},"PeriodicalIF":3.7,"publicationDate":"2025-11-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145520861","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

A multilayered deep learning framework for cyber attack detection and mitigation in a heterogeneous IIoT ecosystem 在异构IIoT生态系统中用于网络攻击检测和缓解的多层深度学习框架

IF 3.7 2区计算机科学 Q2 COMPUTER SCIENCE, INFORMATION SYSTEMS

Journal of Information Security and Applications

Pub Date : 2025-11-13 DOI: 10.1016/j.jisa.2025.104301

Arshad Iqbal, Sohail Asghar, Manzoor Ilahi Tamimy

Intrusion Detection Systems (IDSs) for the Internet of Things (IoT) and Industrial IoT (IIoT) face significant challenges, including high false-positive rates (especially for minority-class attacks) and excessive computational requirements, which hinder their deployment on edge devices. Consequently, alert overload is common because operators receive a large volume of alerts that provide little insight into the problems they address. To address this crucial gap, this study presents DeepGuard, a new four-layer framework that significantly improves the security posture of IoT and industrial IoT environments.

DeepGuard combines binary and multiclass classifications, intelligent alarming, and cyber deception into a single, effective defence mechanism. The system incorporates a random forest classifier for feature selection, which extracts the most relevant data features and processes them for use with an optimised multilayer perceptron (MLP). This method achieved an unprecedented accuracy of 99.9% with a low false-positive rate (FPR) of 0.2%, surpassing the state-of-the-art research studies.

We further demonstrated the practical feasibility of DeepGuard by implementing it on computationally constrained, edge devices. With a computational complexity of

O (n l o g n)

and a memory footprint of less than 100 KB, DeepGuard breaks the long-standing trade-off between detection accuracy and operational performance that has inhibited the adoption of IDS at an industrial scale. In addition to a detection-only approach, DeepGuard includes an embedded honeypot layer that proactively profiles emerging and unknown attacks, thereby enabling automated mitigation responses. Thorough evaluations of the WUSTL-IIoT-2021 and X-IIoTID-2022 datasets demonstrated a new state-of-the-art performance and the feasibility of DeepGuard for protecting critical infrastructure.

物联网（IoT）和工业物联网（IIoT）的入侵检测系统（ids）面临着重大挑战，包括高误报率（特别是针对少数类攻击）和过多的计算需求，这阻碍了它们在边缘设备上的部署。因此，警报过载很常见，因为操作人员接收到大量警报，而这些警报对他们所处理的问题几乎没有提供什么见解。为了解决这一关键差距，本研究提出了DeepGuard，这是一个新的四层框架，可显着改善物联网和工业物联网环境的安全状况。DeepGuard将二进制和多类分类、智能报警和网络欺骗结合到一个单一、有效的防御机制中。该系统采用随机森林分类器进行特征选择，提取最相关的数据特征，并对其进行处理，以便与优化的多层感知器（MLP）一起使用。该方法达到了前所未有的99.9%的准确率和0.2%的低假阳性率（FPR），超过了目前最先进的研究。我们通过在计算受限的边缘设备上实现DeepGuard进一步证明了它的实际可行性。DeepGuard的计算复杂度为0 (nlogn)，内存占用小于100 KB，打破了长期以来在检测精度和操作性能之间的权衡，这种权衡阻碍了IDS在工业规模上的应用。除了仅用于检测的方法外，DeepGuard还包含一个嵌入式蜜罐层，可主动分析新出现的和未知的攻击，从而实现自动缓解响应。对WUSTL-IIoT-2021和X-IIoTID-2022数据集的全面评估证明了DeepGuard在保护关键基础设施方面的最新性能和可行性。

{"title":"A multilayered deep learning framework for cyber attack detection and mitigation in a heterogeneous IIoT ecosystem","authors":"Arshad Iqbal, Sohail Asghar, Manzoor Ilahi Tamimy","doi":"10.1016/j.jisa.2025.104301","DOIUrl":"10.1016/j.jisa.2025.104301","url":null,"abstract":"<div><div>Intrusion Detection Systems (IDSs) for the Internet of Things (IoT) and Industrial IoT (IIoT) face significant challenges, including high false-positive rates (especially for minority-class attacks) and excessive computational requirements, which hinder their deployment on edge devices. Consequently, alert overload is common because operators receive a large volume of alerts that provide little insight into the problems they address. To address this crucial gap, this study presents DeepGuard, a new four-layer framework that significantly improves the security posture of IoT and industrial IoT environments.</div><div>DeepGuard combines binary and multiclass classifications, intelligent alarming, and cyber deception into a single, effective defence mechanism. The system incorporates a random forest classifier for feature selection, which extracts the most relevant data features and processes them for use with an optimised multilayer perceptron (MLP). This method achieved an unprecedented accuracy of 99.9% with a low false-positive rate (FPR) of 0.2%, surpassing the state-of-the-art research studies.</div><div>We further demonstrated the practical feasibility of DeepGuard by implementing it on computationally constrained, edge devices. With a computational complexity of <span><math><mrow><mi>O</mi><mrow><mo>(</mo><mi>n</mi><mi>l</mi><mi>o</mi><mi>g</mi><mi>n</mi><mo>)</mo></mrow></mrow></math></span> and a memory footprint of less than 100 KB, DeepGuard breaks the long-standing trade-off between detection accuracy and operational performance that has inhibited the adoption of IDS at an industrial scale. In addition to a detection-only approach, DeepGuard includes an embedded honeypot layer that proactively profiles emerging and unknown attacks, thereby enabling automated mitigation responses. Thorough evaluations of the WUSTL-IIoT-2021 and X-IIoTID-2022 datasets demonstrated a new state-of-the-art performance and the feasibility of DeepGuard for protecting critical infrastructure.</div></div>","PeriodicalId":48638,"journal":{"name":"Journal of Information Security and Applications","volume":"96 ","pages":"Article 104301"},"PeriodicalIF":3.7,"publicationDate":"2025-11-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145520859","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Beyond Reinforcement Learning for network security: A comprehensive survey and tutorial 超越强化学习的网络安全：一个全面的调查和教程

IF 3.7 2区计算机科学 Q2 COMPUTER SCIENCE, INFORMATION SYSTEMS

Journal of Information Security and Applications

Pub Date : 2025-11-12 DOI: 10.1016/j.jisa.2025.104294

Amir Javadpour , Forough Ja’fari , Tarik Taleb , Fatih Turkmen , Chafika Benzaïd

Maintaining strong security is a complex yet vital challenge in the rapidly evolving landscape of modern digital networks. The risks and consequences of security breaches make neglecting network protection unacceptable. Fortunately, ongoing advances in computer science have equipped researchers with powerful tools to reinforce network defenses. Among these, Reinforcement Learning (RL), a branch of machine learning, has gained significant attention for its versatility and effectiveness in strengthening security mechanisms. This paper presents a comprehensive survey and tutorial on the role of RL in network security. It provides background information, a step-by-step tutorial for training RL models, and systematically categorizes research efforts based on the targeted cyber threats. Leveraging recent advances and real-world applications, this survey elucidates how RL enables the development of adaptive and intelligent systems that autonomously learn and respond to evolving threats. Through in-depth analysis, we provide a comprehensive view of the current landscape and the future potential of RL in safeguarding digital assets. The main contributions of this survey are: (1) a systematic and up-to-date review of RL approaches for network security; (2) a unified taxonomy for classifying RL-based solutions; (3) a comparison of the latest advances from 2019 to 2024 across mainstream and emerging research areas; (4) identification of open challenges and future research directions; and (5) a comparative analysis of state-of-the-art models, offering practical insights for both researchers and practitioners. Furthermore, this survey emphasizes the practical translation of RL advances into real-world deployments. By focusing on hands-on implementation guidelines and comparative analyses of deployment scenarios, it bridges the gap between academic research and operational practice. The comprehensive evaluation of RL-based models across different network environments provides actionable insights for practitioners seeking adaptive and scalable security solutions in dynamic and heterogeneous settings.

在快速发展的现代数字网络环境中，保持强大的安全性是一项复杂而又至关重要的挑战。安全漏洞的风险和后果使得忽视网络保护是不可接受的。幸运的是，计算机科学的不断进步为研究人员提供了强大的工具来加强网络防御。其中，强化学习（RL）作为机器学习的一个分支，因其在加强安全机制方面的多功能性和有效性而受到广泛关注。本文对RL在网络安全中的作用进行了全面的综述和介绍。它提供了背景信息、训练强化学习模型的分步教程，并根据目标网络威胁系统地对研究工作进行了分类。利用最新的进展和现实世界的应用，本调查阐明了强化学习如何使自适应和智能系统的开发能够自主学习和响应不断变化的威胁。通过深入分析，我们全面了解了RL在保护数字资产方面的现状和未来潜力。本调查的主要贡献是：(1)对网络安全的RL方法进行了系统和最新的回顾；(2)基于rl的解决方案的统一分类；(3) 2019 - 2024年主流与新兴研究领域的最新进展对比；(4)确定开放性挑战和未来研究方向；(5)对最先进的模型进行了比较分析，为研究人员和实践者提供了实践见解。此外，本调查强调了将强化学习的进步实际转化为现实世界的部署。通过关注实际的实现指南和部署场景的比较分析，它弥合了学术研究和操作实践之间的差距。跨不同网络环境的基于rl的模型的综合评估为从业者在动态和异构设置中寻求自适应和可扩展的安全解决方案提供了可操作的见解。

{"title":"Beyond Reinforcement Learning for network security: A comprehensive survey and tutorial","authors":"Amir Javadpour , Forough Ja’fari , Tarik Taleb , Fatih Turkmen , Chafika Benzaïd","doi":"10.1016/j.jisa.2025.104294","DOIUrl":"10.1016/j.jisa.2025.104294","url":null,"abstract":"<div><div>Maintaining strong security is a complex yet vital challenge in the rapidly evolving landscape of modern digital networks. The risks and consequences of security breaches make neglecting network protection unacceptable. Fortunately, ongoing advances in computer science have equipped researchers with powerful tools to reinforce network defenses. Among these, Reinforcement Learning (RL), a branch of machine learning, has gained significant attention for its versatility and effectiveness in strengthening security mechanisms. This paper presents a comprehensive survey and tutorial on the role of RL in network security. It provides background information, a step-by-step tutorial for training RL models, and systematically categorizes research efforts based on the targeted cyber threats. Leveraging recent advances and real-world applications, this survey elucidates how RL enables the development of adaptive and intelligent systems that autonomously learn and respond to evolving threats. Through in-depth analysis, we provide a comprehensive view of the current landscape and the future potential of RL in safeguarding digital assets. The main contributions of this survey are: (1) a systematic and up-to-date review of RL approaches for network security; (2) a unified taxonomy for classifying RL-based solutions; (3) a comparison of the latest advances from 2019 to 2024 across mainstream and emerging research areas; (4) identification of open challenges and future research directions; and (5) a comparative analysis of state-of-the-art models, offering practical insights for both researchers and practitioners. Furthermore, this survey emphasizes the practical translation of RL advances into real-world deployments. By focusing on hands-on implementation guidelines and comparative analyses of deployment scenarios, it bridges the gap between academic research and operational practice. The comprehensive evaluation of RL-based models across different network environments provides actionable insights for practitioners seeking adaptive and scalable security solutions in dynamic and heterogeneous settings.</div></div>","PeriodicalId":48638,"journal":{"name":"Journal of Information Security and Applications","volume":"96 ","pages":"Article 104294"},"PeriodicalIF":3.7,"publicationDate":"2025-11-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145520862","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Vulnerabilities in Machine Learning for cybersecurity: Current trends and future research directions 面向网络安全的机器学习漏洞：当前趋势和未来研究方向

IF 3.7 2区计算机科学 Q2 COMPUTER SCIENCE, INFORMATION SYSTEMS

Journal of Information Security and Applications

Pub Date : 2025-11-12 DOI: 10.1016/j.jisa.2025.104269

Shantanu Pal , Geeta Yadav , Zahra Jadidi , Ahsan Habib , Md. Palash Uddin , Chandan Karmakar , Sandeep Shukla

Machine learning (ML) has become integral to cybersecurity applications, e.g., phishing detection, intrusion detection systems, malware analysis, and botnet identification. However, the integration of ML also exposes novel attack surfaces that can be exploited through adversarial machine learning (AML). While prior surveys have examined individual threats or defenses, they often focus narrowly on specific stages, e.g., training or testing. In contrast, in this paper, we provide the first comprehensive survey of adversarial attacks and defenses across the entire ML development life cycle within the cybersecurity domain. Using a structured methodology, we categorize vulnerabilities and countermeasures at each stage, data gathering, model training, testing, deployment, and maintenance, highlighting cross-stage interactions and emerging distributed threat models. Our study addresses key gaps in current defenses, including their limited generalizability and lack of standardized evaluation practices, and identifies promising directions, e.g., lifecycle-aware robustness, distributed resilience, and the integration of statistical with generative methods. Consolidating fragmented research into an end-to-end perspective, this study advances the understanding of AML in cybersecurity and outlines a roadmap for building more trustworthy, and resilient ML-driven security systems.

机器学习（ML）已经成为网络安全应用中不可或缺的一部分，例如网络钓鱼检测、入侵检测系统、恶意软件分析和僵尸网络识别。然而，机器学习的集成也暴露了新的攻击面，可以通过对抗性机器学习（AML）加以利用。虽然之前的调查已经检查了单个威胁或防御，但它们通常只关注特定阶段，例如培训或测试。相比之下，在本文中，我们首次全面调查了网络安全领域内整个机器学习开发生命周期中的对抗性攻击和防御。使用结构化方法，我们对每个阶段的漏洞和对策进行了分类，数据收集，模型训练，测试，部署和维护，突出了跨阶段的交互和新兴的分布式威胁模型。我们的研究解决了当前防御中的关键差距，包括其有限的通用性和缺乏标准化的评估实践，并确定了有前途的方向，例如，生命周期感知的鲁棒性，分布式弹性以及统计与生成方法的集成。本研究将零散的研究整合为端到端视角，促进了对网络安全中的“反洗钱”的理解，并概述了构建更值得信赖、更有弹性的机器学习驱动的安全系统的路线图。

{"title":"Vulnerabilities in Machine Learning for cybersecurity: Current trends and future research directions","authors":"Shantanu Pal , Geeta Yadav , Zahra Jadidi , Ahsan Habib , Md. Palash Uddin , Chandan Karmakar , Sandeep Shukla","doi":"10.1016/j.jisa.2025.104269","DOIUrl":"10.1016/j.jisa.2025.104269","url":null,"abstract":"<div><div>Machine learning (ML) has become integral to cybersecurity applications, e.g., phishing detection, intrusion detection systems, malware analysis, and botnet identification. However, the integration of ML also exposes novel attack surfaces that can be exploited through adversarial machine learning (AML). While prior surveys have examined individual threats or defenses, they often focus narrowly on specific stages, e.g., training or testing. In contrast, in this paper, we provide the first comprehensive survey of adversarial attacks and defenses across the entire ML development life cycle within the cybersecurity domain. Using a structured methodology, we categorize vulnerabilities and countermeasures at each stage, data gathering, model training, testing, deployment, and maintenance, highlighting cross-stage interactions and emerging distributed threat models. Our study addresses key gaps in current defenses, including their limited generalizability and lack of standardized evaluation practices, and identifies promising directions, e.g., lifecycle-aware robustness, distributed resilience, and the integration of statistical with generative methods. Consolidating fragmented research into an end-to-end perspective, this study advances the understanding of AML in cybersecurity and outlines a roadmap for building more trustworthy, and resilient ML-driven security systems.</div></div>","PeriodicalId":48638,"journal":{"name":"Journal of Information Security and Applications","volume":"96 ","pages":"Article 104269"},"PeriodicalIF":3.7,"publicationDate":"2025-11-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145520860","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Blockchain-based access control model for smart grids using peak hour and privilege level attributes (BACS-HP) 基于区块链的智能电网峰值小时和特权级别属性访问控制模型（bac - hp）

IF 3.7 2区计算机科学 Q2 COMPUTER SCIENCE, INFORMATION SYSTEMS

Journal of Information Security and Applications

Pub Date : 2025-11-10 DOI: 10.1016/j.jisa.2025.104261

Sarra Namane , Imed Ben Dhaou

The increasing reliance on smart plugs and smart meters in modern electricity grids introduces significant security vulnerabilities, as unauthorized access can compromise grid reliability and stability. Traditional access control models are ill-suited for smart grids’ decentralized and dynamic nature. This paper introduces BACS-HP, a novel Blockchain-Based Access Control Model for Smart Grids that enhances security by incorporating privilege levels and peak hour attributes. Privilege levels prioritize access to critical devices during energy constraints, while the peak hour attribute enables adaptive decision-making to optimize energy allocation during periods of high demand. Unlike existing blockchain-based access control solutions, BACS-HP uniquely combines these context-aware attributes to provide fine-grained access control tailored to the specific needs of smart grids. The model leverages blockchain technology to ensure the secure and decentralized storage of access rights and enforces policies via smart contracts, mitigating single points of failure. Empirical results demonstrate that BACS-HP achieves low-latency security rule updates (between 42 ms and 46 ms), rapid access request processing (between 21 ms and 46 ms), and a high acceptance rate (60%) for critical devices during power outages, outperforming standard ABAC implementations in terms of responsiveness and prioritization. BACS-HP contributes to advancing access control mechanisms in smart grids and highlights the potential of blockchain to meet the security and performance demands of modern energy systems.

现代电网对智能插头和智能电表的依赖日益增加，这带来了重大的安全漏洞，因为未经授权的访问可能会损害电网的可靠性和稳定性。传统的访问控制模型不适用于智能电网的分散性和动态性。本文介绍了BACS-HP，这是一种新型的基于区块链的智能电网访问控制模型，通过结合特权级别和高峰时间属性来提高安全性。在能源限制期间，特权级别优先考虑对关键设备的访问，而高峰时间属性使自适应决策能够在高需求期间优化能源分配。与现有的基于区块链的访问控制解决方案不同，BACS-HP独特地结合了这些上下文感知属性，提供针对智能电网特定需求的细粒度访问控制。该模型利用区块链技术确保访问权限的安全和分散存储，并通过智能合约执行策略，减少单点故障。实证结果表明，BACS-HP实现了低延迟的安全规则更新（在42 ms到46 ms之间），快速的访问请求处理（在21 ms到46 ms之间），以及在断电期间关键设备的高接受率（60%），在响应性和优先级方面优于标准ABAC实现。BACS-HP有助于推进智能电网中的访问控制机制，并突出区块链在满足现代能源系统安全和性能需求方面的潜力。

{"title":"Blockchain-based access control model for smart grids using peak hour and privilege level attributes (BACS-HP)","authors":"Sarra Namane , Imed Ben Dhaou","doi":"10.1016/j.jisa.2025.104261","DOIUrl":"10.1016/j.jisa.2025.104261","url":null,"abstract":"<div><div>The increasing reliance on smart plugs and smart meters in modern electricity grids introduces significant security vulnerabilities, as unauthorized access can compromise grid reliability and stability. Traditional access control models are ill-suited for smart grids’ decentralized and dynamic nature. This paper introduces BACS-HP, a novel Blockchain-Based Access Control Model for Smart Grids that enhances security by incorporating <em>privilege levels</em> and <em>peak hour</em> attributes. Privilege levels prioritize access to critical devices during energy constraints, while the peak hour attribute enables adaptive decision-making to optimize energy allocation during periods of high demand. Unlike existing blockchain-based access control solutions, BACS-HP uniquely combines these context-aware attributes to provide fine-grained access control tailored to the specific needs of smart grids. The model leverages blockchain technology to ensure the secure and decentralized storage of access rights and enforces policies via smart contracts, mitigating single points of failure. Empirical results demonstrate that BACS-HP achieves low-latency security rule updates (between 42 ms and 46 ms), rapid access request processing (between 21 ms and 46 ms), and a high acceptance rate (60%) for critical devices during power outages, outperforming standard ABAC implementations in terms of responsiveness and prioritization. BACS-HP contributes to advancing access control mechanisms in smart grids and highlights the potential of blockchain to meet the security and performance demands of modern energy systems.</div></div>","PeriodicalId":48638,"journal":{"name":"Journal of Information Security and Applications","volume":"96 ","pages":"Article 104261"},"PeriodicalIF":3.7,"publicationDate":"2025-11-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145520877","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Protocol design of non-linear function in secure multi-party computation based on secret sharing 基于秘密共享的安全多方计算非线性函数协议设计

IF 3.7 2区计算机科学 Q2 COMPUTER SCIENCE, INFORMATION SYSTEMS

Journal of Information Security and Applications

Pub Date : 2025-11-10 DOI: 10.1016/j.jisa.2025.104293

Zhongkai Li, Shuyang Fan, Lingfei Jin

Secure Multi-Party Computation (MPC) enables a group of untrusted parties to collaboratively compute the output of a specified function, while ensuring that each party’s private input remains confidential. Coupled with secret sharing, MPC facilitates privacy-preserving computations, a technique increasingly utilized in diverse fields, such as machine learning. While efficient protocols exist within MPC for linear functions, the evaluation of non-linear functions presents a significant challenge. Existing methods for non-linear functions are often either inefficient or lack the generality for widespread adoption, making them a major impediment in both the design and practical implementation of MPC schemes.

In this study, we explore the development of a generic protocol for non-linear function computation in MPC, grounded in secret sharing. We have devised a series of protocols to compute fundamental non-linear functions in a three-party setting under a semi-honest security model, representing secret-shared decimal numbers in fixed-point format. These protocols include

Π_{exp}

for exponential functions,

Π_{log}

for logarithmic functions, and

Π_{Inv}

for inverse proportion functions. By integrating these basic functions, we can formulate protocols for a broad spectrum of non-linear functions. Specifically, we have developed the

Π_{Sigmoid}

and

Π_{Tanh}

protocols based on the aforementioned methods. Throughout this paper, unless otherwise specified, comparisons refer exclusively to secret-sharing-based (SS-based) MPC protocols in the three-party, semi-honest setting; constant-round garbled-circuit (GC) approaches are outside our comparison scope due to different cost trade-offs. Within this SS-based literature, our protocols offer the lowest online communication rounds. Furthermore,

Π_{exp}

and

Π_{inv}

support an extended range of inputs, and

Π_{log}

represents the first protocol capable of handling logarithmic functions with fixed-point inputs. This paper provides a thorough analysis of the security and performance of these innovative protocols.

安全多方计算（MPC）使一组不受信任的各方能够协作计算指定函数的输出，同时确保每一方的私有输入保持机密。再加上秘密共享，MPC促进了隐私保护计算，这是一种越来越多地应用于不同领域的技术，如机器学习。虽然MPC中存在用于线性函数的有效协议，但非线性函数的评估提出了一个重大挑战。非线性函数的现有方法往往效率低下或缺乏广泛采用的通用性，使它们成为MPC方案设计和实际实施的主要障碍。在本研究中，我们探索了基于秘密共享的MPC非线性函数计算通用协议的开发。我们设计了一系列协议，在半诚实的安全模型下计算三方设置中的基本非线性函数，以定点格式表示秘密共享的十进制数。这些协议包括用于指数函数的Πexp，用于对数函数的Πlog和用于反比函数的ΠInv。通过整合这些基本函数，我们可以为广泛的非线性函数制定协议。具体来说，我们基于上述方法开发了ΠSigmoid和ΠTanh协议。在本文中，除非另有说明，比较只指在三方、半诚实设置中基于秘密共享（SS-based）的MPC协议；由于不同的成本权衡，恒圆乱码电路（GC）方法超出了我们的比较范围。在这个基于ss的文献中，我们的协议提供了最低的在线通信回合。此外，Πexp和Πinv支持扩展的输入范围，Πlog代表了第一个能够处理具有定点输入的对数函数的协议。本文对这些创新协议的安全性和性能进行了全面的分析。

{"title":"Protocol design of non-linear function in secure multi-party computation based on secret sharing","authors":"Zhongkai Li, Shuyang Fan, Lingfei Jin","doi":"10.1016/j.jisa.2025.104293","DOIUrl":"10.1016/j.jisa.2025.104293","url":null,"abstract":"<div><div>Secure Multi-Party Computation (MPC) enables a group of untrusted parties to collaboratively compute the output of a specified function, while ensuring that each party’s private input remains confidential. Coupled with secret sharing, MPC facilitates privacy-preserving computations, a technique increasingly utilized in diverse fields, such as machine learning. While efficient protocols exist within MPC for linear functions, the evaluation of non-linear functions presents a significant challenge. Existing methods for non-linear functions are often either inefficient or lack the generality for widespread adoption, making them a major impediment in both the design and practical implementation of MPC schemes.</div><div>In this study, we explore the development of a generic protocol for non-linear function computation in MPC, grounded in secret sharing. We have devised a series of protocols to compute fundamental non-linear functions in a three-party setting under a semi-honest security model, representing secret-shared decimal numbers in fixed-point format. These protocols include <span><math><msub><mrow><mi>Π</mi></mrow><mrow><mo>exp</mo></mrow></msub></math></span> for exponential functions, <span><math><msub><mrow><mi>Π</mi></mrow><mrow><mo>log</mo></mrow></msub></math></span> for logarithmic functions, and <span><math><msub><mrow><mi>Π</mi></mrow><mrow><mtext>Inv</mtext></mrow></msub></math></span> for inverse proportion functions. By integrating these basic functions, we can formulate protocols for a broad spectrum of non-linear functions. Specifically, we have developed the <span><math><msub><mrow><mi>Π</mi></mrow><mrow><mtext>Sigmoid</mtext></mrow></msub></math></span> and <span><math><msub><mrow><mi>Π</mi></mrow><mrow><mtext>Tanh</mtext></mrow></msub></math></span> protocols based on the aforementioned methods. Throughout this paper, unless otherwise specified, comparisons refer exclusively to secret-sharing-based (SS-based) MPC protocols in the three-party, semi-honest setting; constant-round garbled-circuit (GC) approaches are outside our comparison scope due to different cost trade-offs. Within this SS-based literature, our protocols offer the lowest online communication rounds. Furthermore, <span><math><msub><mrow><mi>Π</mi></mrow><mrow><mo>exp</mo></mrow></msub></math></span> and <span><math><msub><mrow><mi>Π</mi></mrow><mrow><mtext>inv</mtext></mrow></msub></math></span> support an extended range of inputs, and <span><math><msub><mrow><mi>Π</mi></mrow><mrow><mo>log</mo></mrow></msub></math></span> represents the first protocol capable of handling logarithmic functions with fixed-point inputs. This paper provides a thorough analysis of the security and performance of these innovative protocols.</div></div>","PeriodicalId":48638,"journal":{"name":"Journal of Information Security and Applications","volume":"96 ","pages":"Article 104293"},"PeriodicalIF":3.7,"publicationDate":"2025-11-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145520879","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Uncertainty-aware regular-singular discriminant analysis for lossless watermarking 无损水印的正则奇异判别分析

IF 3.7 2区计算机科学 Q2 COMPUTER SCIENCE, INFORMATION SYSTEMS

Journal of Information Security and Applications

Pub Date : 2025-11-08 DOI: 10.1016/j.jisa.2025.104295

Guo-Dong Su , Xu Wang , Ching-Chun Chang

It remains a major challenge in how to effectively organize and manage digital images stored in cloud. Regular-singular (RS) based watermarking, as one of important technologies, aims to insert watermark into digital images to solve this issue. By revisiting series of RS based watermarking methods, however, how to achieve a better trade-off between enlarging the embedding capacity and keeping the amount of distortion as soon as possible remains an interesting problem, especially deep learning comes powerful. For this, this paper presents a novel lossless watermarking method using uncertainty-aware discriminant analysis and deep learning technology. First, a numerical ordinary differential equation inspired network architecture for cover synthesis we refer to as NDCS is introduced. It produces a more realistic cover objective by minimizing a smaller local truncation error. As for NDCS, we are also interested in its performance under different network configurations. On this basis, we introduce an uncertainty-aware discriminant analysis in steganographic algorithm, thereby enabling to yield perceptually indistinguishable watermarked images at various capacities. The experimental results demonstrate that our method is conducive to improving the quality of synthetic objective with the mean hamming distance of 0.2231 and achieving a more satisfactory rate-distortion trade-off with an average embedding capacity of 0.2043 bpp, when comparing to the prior regular-singular methods. In addition, our approach can against RS steganalysis and has the identical performance in encrypted domain.

如何有效地组织和管理存储在云中的数字图像仍然是一个重大挑战。正则奇异水印作为一种重要的水印技术，旨在通过在数字图像中插入水印来解决这一问题。然而，通过回顾一系列基于RS的水印方法，如何在扩大嵌入容量和尽快保持失真量之间取得更好的平衡仍然是一个有趣的问题，尤其是深度学习的强大。为此，本文提出了一种利用不确定性感知判别分析和深度学习技术的无损水印方法。首先，介绍了一种数值常微分方程启发的覆盖物综合网络结构，我们称之为NDCS。它通过最小化较小的局部截断误差来产生更真实的覆盖目标。对于NDCS，我们也对它在不同网络配置下的性能很感兴趣。在此基础上，我们在隐写算法中引入了不确定性感知的判别分析，从而能够在不同容量下产生感知上不可区分的水印图像。实验结果表明，与现有的正则奇异方法相比，该方法可以提高合成目标的质量，平均汉明距离为0.2231，平均嵌入容量为0.2043 bpp，实现了更令人满意的率失真权衡。此外，我们的方法可以对抗RS隐写分析，并且在加密域具有相同的性能。

{"title":"Uncertainty-aware regular-singular discriminant analysis for lossless watermarking","authors":"Guo-Dong Su , Xu Wang , Ching-Chun Chang","doi":"10.1016/j.jisa.2025.104295","DOIUrl":"10.1016/j.jisa.2025.104295","url":null,"abstract":"<div><div>It remains a major challenge in how to effectively organize and manage digital images stored in cloud. Regular-singular (RS) based watermarking, as one of important technologies, aims to insert watermark into digital images to solve this issue. By revisiting series of RS based watermarking methods, however, how to achieve a better trade-off between enlarging the embedding capacity and keeping the amount of distortion as soon as possible remains an interesting problem, especially deep learning comes powerful. For this, this paper presents a novel lossless watermarking method using uncertainty-aware discriminant analysis and deep learning technology. First, a numerical ordinary differential equation inspired network architecture for cover synthesis we refer to as <strong>NDCS</strong> is introduced. It produces a more realistic cover objective by minimizing a smaller local truncation error. As for <strong>NDCS</strong>, we are also interested in its performance under different network configurations. On this basis, we introduce an uncertainty-aware discriminant analysis in steganographic algorithm, thereby enabling to yield perceptually indistinguishable watermarked images at various capacities. The experimental results demonstrate that our method is conducive to improving the quality of synthetic objective with the mean hamming distance of 0.2231 and achieving a more satisfactory rate-distortion trade-off with an average embedding capacity of 0.2043 bpp, when comparing to the prior regular-singular methods. In addition, our approach can against RS steganalysis and has the identical performance in encrypted domain.</div></div>","PeriodicalId":48638,"journal":{"name":"Journal of Information Security and Applications","volume":"96 ","pages":"Article 104295"},"PeriodicalIF":3.7,"publicationDate":"2025-11-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145468934","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

LeakyDroid: A lightweight method for detecting zero-day leaky Android applications using One-Class Graph Neural Networks LeakyDroid：使用一类图神经网络检测零日漏洞的Android应用程序的轻量级方法

IF 3.7 2区计算机科学 Q2 COMPUTER SCIENCE, INFORMATION SYSTEMS

Journal of Information Security and Applications

Pub Date : 2025-11-07 DOI: 10.1016/j.jisa.2025.104296

Neha Sharma, Mayank Swarnkar, Shaan Kumar

In the current era of mobile technology, ensuring user data security and privacy is very important, particularly with the rise of malicious Android applications that aim to leak end-user data. Moreover, the popularity of the Android OS is resulting in growing numbers of such malicious Android applications. Hackers make the apps malicious by downloading the source code of Android applications and modifying it. Static analysis techniques have traditionally been used to detect such leaky Android applications. However, these methods cannot simulate runtime behaviours, leading to false positives or negatives. Moreover, obfuscated code is also harder to analyse using this technique. On the other hand, dynamic analysis-based methods are used to overcome these issues because they capture the application’s actual behaviour during runtime. However, dynamic analysis methods have high computational complexity. To fill this gap, we propose LeakyDroid, a static but lightweight method for detecting zero-day leaky Android applications using one-class graph neural networks. LeakyDroid distinguishes between the zero-day malicious and genuine versions of Android applications based on function calls inside various class files of the installable APK files. LeakyDroid generates a control flow graph from function calls from several versions of normal APK files of the same application. The graph is trained using OCGNN, which effectively captures relationships and invocation patterns of normal APK files. While testing an unknown version of the same application’s APK, if a considerable deviation is seen from normal behaviour, the application is detected as malicious. We evaluated the performance of LeakyDroid on three applications, namely WhatsApp, Netflix, and Instagram, each with approximately 25 benign and a few malicious and leaky versions. LeakyDroid successfully detected all the malicious versions of APK with no false positives.

在当前的移动技术时代，确保用户数据的安全和隐私是非常重要的，特别是随着旨在泄露最终用户数据的恶意Android应用程序的兴起。此外，Android操作系统的普及导致了越来越多的恶意Android应用程序。黑客通过下载Android应用程序的源代码并对其进行修改，使这些应用程序具有恶意。传统上，静态分析技术被用于检测此类泄漏的Android应用程序。然而，这些方法不能模拟运行时行为，从而导致误报或误报。此外，使用这种技术分析混淆的代码也更加困难。另一方面，基于动态分析的方法用于克服这些问题，因为它们在运行时捕获应用程序的实际行为。然而，动态分析方法具有较高的计算复杂度。为了填补这一空白，我们提出了LeakyDroid，这是一种静态但轻量级的方法，用于使用一类图神经网络检测零日漏洞的Android应用程序。LeakyDroid根据可安装APK文件的各种类文件中的函数调用来区分零日恶意版本和正版Android应用程序。LeakyDroid从来自同一应用程序的几个版本的普通APK文件的函数调用中生成控制流图。该图使用OCGNN进行训练，OCGNN有效地捕获普通APK文件的关系和调用模式。在测试同一应用程序APK的未知版本时，如果看到与正常行为有相当大的偏差，则检测到该应用程序为恶意应用程序。我们评估了LeakyDroid在三个应用程序上的表现，分别是WhatsApp、Netflix和Instagram，每个应用程序都有大约25个良性版本和一些恶意和泄漏版本。LeakyDroid成功检测到所有恶意版本的APK，没有误报。

{"title":"LeakyDroid: A lightweight method for detecting zero-day leaky Android applications using One-Class Graph Neural Networks","authors":"Neha Sharma, Mayank Swarnkar, Shaan Kumar","doi":"10.1016/j.jisa.2025.104296","DOIUrl":"10.1016/j.jisa.2025.104296","url":null,"abstract":"<div><div>In the current era of mobile technology, ensuring user data security and privacy is very important, particularly with the rise of malicious Android applications that aim to leak end-user data. Moreover, the popularity of the Android OS is resulting in growing numbers of such malicious Android applications. Hackers make the apps malicious by downloading the source code of Android applications and modifying it. Static analysis techniques have traditionally been used to detect such leaky Android applications. However, these methods cannot simulate runtime behaviours, leading to false positives or negatives. Moreover, obfuscated code is also harder to analyse using this technique. On the other hand, dynamic analysis-based methods are used to overcome these issues because they capture the application’s actual behaviour during runtime. However, dynamic analysis methods have high computational complexity. To fill this gap, we propose <em>LeakyDroid</em>, a static but lightweight method for detecting zero-day leaky Android applications using one-class graph neural networks. <em>LeakyDroid</em> distinguishes between the zero-day malicious and genuine versions of Android applications based on function calls inside various class files of the installable APK files. <em>LeakyDroid</em> generates a control flow graph from function calls from several versions of normal APK files of the same application. The graph is trained using OCGNN, which effectively captures relationships and invocation patterns of normal APK files. While testing an unknown version of the same application’s APK, if a considerable deviation is seen from normal behaviour, the application is detected as malicious. We evaluated the performance of <em>LeakyDroid</em> on three applications, namely WhatsApp, Netflix, and Instagram, each with approximately 25 benign and a few malicious and leaky versions. <em>LeakyDroid</em> successfully detected all the malicious versions of APK with no false positives.</div></div>","PeriodicalId":48638,"journal":{"name":"Journal of Information Security and Applications","volume":"96 ","pages":"Article 104296"},"PeriodicalIF":3.7,"publicationDate":"2025-11-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145468937","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0