Information Processing & Management最新文献_第3页

Pontis: A decentralized framework for unifying remote attestation and enabling interoperability between heterogeneous TEEs Pontis：一个分散的框架，用于统一远程认证和支持异构tee之间的互操作性

IF 6.9 1区管理学 Q1 COMPUTER SCIENCE, INFORMATION SYSTEMS

Information Processing & Management

Pub Date : 2026-02-03 DOI: 10.1016/j.ipm.2026.104652

Jun Li , Hong Lei , Xinman Luo , Lei Zhang , Pengxu Shen , Tianyu Huang , Jieren Cheng

In modern decentralized information systems, establishing verifiable trust in remote computing environments has emerged as a critical challenge for secure cross-domain collaboration. Hardware-based Trusted Execution Environments (TEEs) such as Intel SGX and ARM TrustZone offer a promising foundation for addressing this challenge through cryptographically verifiable execution guarantees, but their incompatible Remote Attestation (RA) mechanisms create fundamental barriers to cross-platform trust establishment. Current solutions either focus on single-vendor ecosystems or introduce prohibitive architectural complexity, failing to address the critical need for lightweight, decentralized interoperability. This paper presents Pontis, a decentralized blockchain-based framework that solves unified RA and cross-platform communication challenges for heterogeneous TEEs through three key innovations: (1) a distributed off-chain Coordinator that normalizes vendor-specific attestation protocols, combined with blockchain-anchored decentralized identifiers that provide immutable distributed identities for TEE instances; (2) blockchain-based smart contracts implementing Registration, Attestation, and Management functions for immutable trust status propagation; and (3) secure cross-TEE communication channels built on the Noise protocol framework. Pontis reduces attestation complexity from O(n²) → O(n), achieving up to 99% reduction in required attestations for large-scale deployments. Comprehensive evaluations on heterogeneous TEE platforms, including Intel SGX and ARM TrustZone, demonstrate that Pontis maintains an attestation latency of 60 ms with over 10,000 TEE instances, while establishing a trusted channel between heterogeneous TEEs requires only 5 ms. These results demonstrate the robustness and feasibility of Pontis, establishing a robust foundation for secure, scalable, and flexible cross-platform collaboration in demanding heterogeneous computing environments.

在现代分散的信息系统中，在远程计算环境中建立可验证的信任已成为安全跨域协作的关键挑战。基于硬件的可信执行环境（tee），如Intel SGX和ARM TrustZone，通过加密可验证的执行保证，为解决这一挑战提供了有希望的基础，但它们不兼容的远程认证（RA）机制为跨平台信任的建立创造了根本障碍。当前的解决方案要么专注于单一供应商的生态系统，要么引入了令人望而却步的架构复杂性，未能解决轻量级、分散互操作性的关键需求。本文介绍了Pontis，这是一个基于区块链的去中心化框架，通过三个关键创新解决了异构TEE的统一RA和跨平台通信挑战：(1)分布式链下协调器，用于规范供应商特定的认证协议，结合区块链锚定的去中心化标识符，为TEE实例提供不可变的分布式身份；(2)基于区块链的智能合约，实现不可变信任状态传播的注册、认证和管理功能；(3)建立在噪声协议框架上的安全跨tee通信通道。Pontis将认证复杂度从O（n2） → O(n)降低，可将大规模部署所需的认证降低高达99%。对异构TEE平台（包括Intel SGX和ARM TrustZone）的综合评估表明，Pontis在超过10,000个TEE实例中保持了60毫秒的认证延迟，而在异构TEE之间建立可信通道只需要5毫秒。这些结果证明了Pontis的健壮性和可行性，在要求苛刻的异构计算环境中为安全、可扩展和灵活的跨平台协作建立了坚实的基础。

{"title":"Pontis: A decentralized framework for unifying remote attestation and enabling interoperability between heterogeneous TEEs","authors":"Jun Li , Hong Lei , Xinman Luo , Lei Zhang , Pengxu Shen , Tianyu Huang , Jieren Cheng","doi":"10.1016/j.ipm.2026.104652","DOIUrl":"10.1016/j.ipm.2026.104652","url":null,"abstract":"<div><div>In modern decentralized information systems, establishing verifiable trust in remote computing environments has emerged as a critical challenge for secure cross-domain collaboration. Hardware-based Trusted Execution Environments (TEEs) such as Intel SGX and ARM TrustZone offer a promising foundation for addressing this challenge through cryptographically verifiable execution guarantees, but their incompatible Remote Attestation (RA) mechanisms create fundamental barriers to cross-platform trust establishment. Current solutions either focus on single-vendor ecosystems or introduce prohibitive architectural complexity, failing to address the critical need for lightweight, decentralized interoperability. This paper presents Pontis, a decentralized blockchain-based framework that solves unified RA and cross-platform communication challenges for heterogeneous TEEs through three key innovations: (1) a distributed off-chain Coordinator that normalizes vendor-specific attestation protocols, combined with blockchain-anchored decentralized identifiers that provide immutable distributed identities for TEE instances; (2) blockchain-based smart contracts implementing Registration, Attestation, and Management functions for immutable trust status propagation; and (3) secure cross-TEE communication channels built on the Noise protocol framework. Pontis reduces attestation complexity from <em>O</em>(<em>n</em><sup>2</sup>) → <em>O</em>(<em>n</em>), achieving up to 99% reduction in required attestations for large-scale deployments. Comprehensive evaluations on heterogeneous TEE platforms, including Intel SGX and ARM TrustZone, demonstrate that Pontis maintains an attestation latency of 60 ms with over 10,000 TEE instances, while establishing a trusted channel between heterogeneous TEEs requires only 5 ms. These results demonstrate the robustness and feasibility of Pontis, establishing a robust foundation for secure, scalable, and flexible cross-platform collaboration in demanding heterogeneous computing environments.</div></div>","PeriodicalId":50365,"journal":{"name":"Information Processing & Management","volume":"63 5","pages":"Article 104652"},"PeriodicalIF":6.9,"publicationDate":"2026-02-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146190593","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"管理学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Source-free time series domain adaptation with class-aware temporal imputation 无源时间序列域自适应与类感知时间输入

IF 6.9 1区管理学 Q1 COMPUTER SCIENCE, INFORMATION SYSTEMS

Information Processing & Management

Pub Date : 2026-02-03 DOI: 10.1016/j.ipm.2026.104658

Yingyi Zhong, Wen’an Zhou, Liwen Tao

Source-free domain adaptation (SFDA) aims to adapt a pre-trained source model to the target domain without access to the source domain data. Recent time series SFDA works transfer temporal dynamics by employing a self-supervised temporal imputation approach designed to reconstruct masked sequences. However, this approach suffers from semantic ambiguity in the reconstructed input and task/classification-irrelevance in the reconstruction objective, which leads to the neglect of class-specific temporal details. As a result, the source model struggles to learn sufficiently discriminative temporal dynamics, limiting its performance on the target domain classification task. To address these issues, we propose Class-Aware Temporal Imputation (CATI), a novel time series SFDA method. CATI improves the existing temporal imputation approach during source pre-training by introducing a class-aware mechanism comprising two modules: Task-guided Contrastive Learning (TCL) and Gradient-based Semantic Enhancement (GSE). TCL refines the global semantic structure through contrastive learning to improve the semantic clarity of the reconstructed input, while GSE highlights local task-relevant regions through backpropagated gradients to guide the reconstruction process towards discriminative temporal details. By combining TCL and GSE, CATI enables the source model to better capture class-specific temporal dynamics, leading to improved adaptation performance on the target domain. Extensive experiments on four real-world time series datasets demonstrate that CATI surpasses state-of-the-art domain adaptation methods, with average gains of 1.41% in accuracy and 1.54% in MF1-score.

无源域自适应（SFDA）是指在不访问源域数据的情况下，将预先训练好的源模型适应目标域。最近的时间序列SFDA通过采用旨在重建掩码序列的自监督时间imputation方法来转移时间动力学。然而，这种方法存在重构输入中的语义模糊和重构目标中的任务/分类不相关的问题，从而导致忽略特定于类的时间细节。因此，源模型难以充分学习判别时态动态，限制了其在目标领域分类任务上的性能。为了解决这些问题，我们提出了一种新的时间序列SFDA方法——类别感知时间Imputation （CATI）。CATI通过引入类别感知机制，改进了现有的源预训练过程中的时间imputation方法，该机制包括两个模块：任务引导的对比学习（Task-guided Contrastive Learning， TCL）和基于梯度的语义增强（Gradient-based Semantic Enhancement， GSE）。TCL通过对比学习细化全局语义结构，提高重构输入的语义清晰度；GSE通过反向传播梯度突出局部任务相关区域，引导重构过程走向判别性的时间细节。通过结合TCL和GSE， CATI使源模型能够更好地捕获特定于类的时间动态，从而提高目标域上的自适应性能。在四个真实时间序列数据集上的大量实验表明，CATI优于最先进的领域自适应方法，准确率平均提高1.41%，MF1-score平均提高1.54%。

{"title":"Source-free time series domain adaptation with class-aware temporal imputation","authors":"Yingyi Zhong, Wen’an Zhou, Liwen Tao","doi":"10.1016/j.ipm.2026.104658","DOIUrl":"10.1016/j.ipm.2026.104658","url":null,"abstract":"<div><div>Source-free domain adaptation (SFDA) aims to adapt a pre-trained source model to the target domain without access to the source domain data. Recent time series SFDA works transfer temporal dynamics by employing a self-supervised temporal imputation approach designed to reconstruct masked sequences. However, this approach suffers from semantic ambiguity in the reconstructed input and task/classification-irrelevance in the reconstruction objective, which leads to the neglect of class-specific temporal details. As a result, the source model struggles to learn sufficiently discriminative temporal dynamics, limiting its performance on the target domain classification task. To address these issues, we propose Class-Aware Temporal Imputation (CATI), a novel time series SFDA method. CATI improves the existing temporal imputation approach during source pre-training by introducing a class-aware mechanism comprising two modules: Task-guided Contrastive Learning (TCL) and Gradient-based Semantic Enhancement (GSE). TCL refines the global semantic structure through contrastive learning to improve the semantic clarity of the reconstructed input, while GSE highlights local task-relevant regions through backpropagated gradients to guide the reconstruction process towards discriminative temporal details. By combining TCL and GSE, CATI enables the source model to better capture class-specific temporal dynamics, leading to improved adaptation performance on the target domain. Extensive experiments on four real-world time series datasets demonstrate that CATI surpasses state-of-the-art domain adaptation methods, with average gains of 1.41% in accuracy and 1.54% in MF1-score.</div></div>","PeriodicalId":50365,"journal":{"name":"Information Processing & Management","volume":"63 5","pages":"Article 104658"},"PeriodicalIF":6.9,"publicationDate":"2026-02-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146190591","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"管理学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

FedDPKD: Federated learning with dual-phase knowledge distillation for label distribution skew FedDPKD：基于双阶段知识蒸馏的标签分布倾斜联邦学习

IF 6.9 1区管理学 Q1 COMPUTER SCIENCE, INFORMATION SYSTEMS

Information Processing & Management

Pub Date : 2026-02-03 DOI: 10.1016/j.ipm.2026.104657

Fanfan Shen , Wenzhang Su , Chao Xu , Zhiquan Liu , Jun Feng , Yanxiang He

The non-independent and identically distributed (non-IID) nature of client data presents a major challenge in federated learning, particularly in the form of label distribution skew. This skew often exacerbates the tendency of models to forget minority and absent labels, resulting in classification outcomes biased toward majority labels. To address this issue, we propose a novel approach called Federated Learning with Dual-Phase Knowledge Distillation (FedDPKD). In this framework, client labels are dynamically categorized into majority, minority, and absent labels based on local label distribution. FedDPKD employs a dual-phase knowledge distillation strategy, consisting of Global-to-Client (G2C) and Client-to-Global (C2G) distillation stages, enhanced by a dynamic weight adjustment mechanism to improve training flexibility and efficiency. The proposed method enhances the global model’s classification accuracy and significantly alleviates the forgetting of minority and absent labels under label distribution skew. Extensive experiments on CIFAR-10, CIFAR-100, MNIST, and CINIC-10 demonstrate that FedDPKD consistently outperforms existing leading methods. Notably, under severe label distribution skew, FedDPKD achieves a 6.4% improvement in accuracy over the second-best baseline on the CIFAR-10 dataset.

客户端数据的非独立和同分布（non-IID）特性在联邦学习中提出了一个主要挑战，特别是以标签分布倾斜的形式。这种偏差往往加剧了模型忘记少数和缺失标签的趋势，导致分类结果偏向多数标签。为了解决这个问题，我们提出了一种新的方法，称为双阶段知识蒸馏联邦学习（FedDPKD）。在该框架中，客户端标签根据本地标签分布动态地分为多数标签、少数标签和缺席标签。FedDPKD采用双阶段知识蒸馏策略，包括全球到客户（G2C）和客户到全球（C2G）蒸馏阶段，并通过动态权重调整机制增强，以提高培训的灵活性和效率。该方法提高了全局模型的分类精度，显著缓解了标签分布偏态下的少数标签和缺席标签的遗忘现象。在CIFAR-10、CIFAR-100、MNIST和CINIC-10上的大量实验表明，FedDPKD始终优于现有的领先方法。值得注意的是，在严重的标签分布倾斜下，FedDPKD在CIFAR-10数据集上比第二好的基线精度提高了6.4%。

{"title":"FedDPKD: Federated learning with dual-phase knowledge distillation for label distribution skew","authors":"Fanfan Shen , Wenzhang Su , Chao Xu , Zhiquan Liu , Jun Feng , Yanxiang He","doi":"10.1016/j.ipm.2026.104657","DOIUrl":"10.1016/j.ipm.2026.104657","url":null,"abstract":"<div><div>The non-independent and identically distributed (non-IID) nature of client data presents a major challenge in federated learning, particularly in the form of label distribution skew. This skew often exacerbates the tendency of models to forget minority and absent labels, resulting in classification outcomes biased toward majority labels. To address this issue, we propose a novel approach called Federated Learning with Dual-Phase Knowledge Distillation (FedDPKD). In this framework, client labels are dynamically categorized into majority, minority, and absent labels based on local label distribution. FedDPKD employs a dual-phase knowledge distillation strategy, consisting of Global-to-Client (G2C) and Client-to-Global (C2G) distillation stages, enhanced by a dynamic weight adjustment mechanism to improve training flexibility and efficiency. The proposed method enhances the global model’s classification accuracy and significantly alleviates the forgetting of minority and absent labels under label distribution skew. Extensive experiments on CIFAR-10, CIFAR-100, MNIST, and CINIC-10 demonstrate that FedDPKD consistently outperforms existing leading methods. Notably, under severe label distribution skew, FedDPKD achieves a 6.4% improvement in accuracy over the second-best baseline on the CIFAR-10 dataset.</div></div>","PeriodicalId":50365,"journal":{"name":"Information Processing & Management","volume":"63 5","pages":"Article 104657"},"PeriodicalIF":6.9,"publicationDate":"2026-02-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146190715","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"管理学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

DUIC: User-descriptive intention guided clustering for personalized and understandable document partitions DUIC：用户描述意图引导的用于个性化和可理解文档分区的聚类

IF 6.9 1区管理学 Q1 COMPUTER SCIENCE, INFORMATION SYSTEMS

Information Processing & Management

Pub Date : 2026-02-03 DOI: 10.1016/j.ipm.2026.104648

Jingjing Xue , Ruizhang Huang , Ruina Bai , Yongbin Qin , Yanping Chen , Chuan Lin

Personalized document clustering often fails to directly utilize user-descriptive intentions expressed in natural language and produces clusters that are difficult to understand. We propose Descriptive-Intention-Guided Understandable Clustering (DUIC), an end-to-end model that integrates a Personalized Intention-Guided Clustering (PIGC) module with a User-Aligned Cluster Explanation (UACE) module. PIGC parses and leverages user-descriptive intentions to guide clustering, while UACE generates human-understandable explanations aligned with these intentions. DUIC was evaluated on six benchmark datasets (737–6,000 text documents) including MIND-6K, BBCSport, and ArXiv variants, against ten state-of-the-art baselines. It achieved up to 99.05% accuracy, 92.80% NMI, and 96.24% ARI. Compared with the strongest baseline on each dataset, DUIC achieved maximum improvements of 16.28% in ACC and 10.93% in NMI. Results show DUIC not only effectively guides clustering results to match user intention but also helps users understand what each cluster represents, bridging the gap between user-descriptive intention and personalized, understandable clustering results.

个性化文档聚类通常不能直接利用用自然语言表达的用户描述意图，产生难以理解的聚类。我们提出了描述-意图引导的可理解聚类（DUIC），这是一种集成了个性化意图引导聚类（PIGC）模块和用户对齐聚类解释（UACE）模块的端到端模型。PIGC解析并利用用户描述意图来指导聚类，而UACE生成与这些意图一致的人类可理解的解释。DUIC在六个基准数据集（737 - 6000个文本文档）上进行了评估，包括MIND-6K、BBCSport和ArXiv变体，以及10个最先进的基线。准确率达到99.05%，NMI达到92.80%，ARI达到96.24%。与每个数据集的最强基线相比，DUIC在ACC和NMI方面的最大改善分别为16.28%和10.93%。结果表明，DUIC不仅可以有效地引导聚类结果匹配用户意图，还可以帮助用户理解每个聚类代表什么，弥合了用户描述性意图与个性化、可理解的聚类结果之间的差距。

{"title":"DUIC: User-descriptive intention guided clustering for personalized and understandable document partitions","authors":"Jingjing Xue , Ruizhang Huang , Ruina Bai , Yongbin Qin , Yanping Chen , Chuan Lin","doi":"10.1016/j.ipm.2026.104648","DOIUrl":"10.1016/j.ipm.2026.104648","url":null,"abstract":"<div><div>Personalized document clustering often fails to directly utilize user-descriptive intentions expressed in natural language and produces clusters that are difficult to understand. We propose Descriptive-Intention-Guided Understandable Clustering (DUIC), an end-to-end model that integrates a Personalized Intention-Guided Clustering (PIGC) module with a User-Aligned Cluster Explanation (UACE) module. PIGC parses and leverages user-descriptive intentions to guide clustering, while UACE generates human-understandable explanations aligned with these intentions. DUIC was evaluated on six benchmark datasets (737–6,000 text documents) including MIND-6K, BBCSport, and ArXiv variants, against ten state-of-the-art baselines. It achieved up to 99.05% accuracy, 92.80% NMI, and 96.24% ARI. Compared with the strongest baseline on each dataset, DUIC achieved maximum improvements of 16.28% in ACC and 10.93% in NMI. Results show DUIC not only effectively guides clustering results to match user intention but also helps users understand what each cluster represents, bridging the gap between user-descriptive intention and personalized, understandable clustering results.</div></div>","PeriodicalId":50365,"journal":{"name":"Information Processing & Management","volume":"63 5","pages":"Article 104648"},"PeriodicalIF":6.9,"publicationDate":"2026-02-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146190674","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"管理学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

SageJavon: A scalable AI tutor for personalized programming learning SageJavon：一个可扩展的AI导师，用于个性化编程学习

IF 6.9 1区管理学 Q1 COMPUTER SCIENCE, INFORMATION SYSTEMS

Information Processing & Management

Pub Date : 2026-02-03 DOI: 10.1016/j.ipm.2025.104605

Hanyu Zhao , Yuzhuo Wu , Zhufeng Lu , Xiaohua Yu , Weikai Miao , Liangyu Chen

Despite advancements in programming education, providing adaptive learning experiences while managing cognitive load remains a challenge. Inspired by Bloom’s Taxonomy, we propose the Learn-Practice-Evaluate-Support (LPES) framework, integrated into SageJavon, an AI-powered tutoring system based on Large Language Models (LLMs). SageJavon overcomes the limitations of traditional AI tutors, such as rigid content delivery and a lack of personalized feedback. It includes four innovations: (1) an enhanced Retrieval-Augmented Generation (RAG) model for personalized resource delivery, (2) heuristic and follow-up questions to guide problem-solving, (3) adaptive knowledge tracing and recommendation models, and (4) the FUPS-Score for automatic code assessment. Deployed in a 12-week course with 85 students, SageJavon facilitated 5306 interactions in the Knowledge Q&A module and 5144 in the Programming Mentor module, with LLM-based scoring yielding average scores of 87.39 and 84.19. Compared to traditional tools, SageJavon reduces cognitive load, showing improved scores in physical demand, temporal demand, and performance (all p < .001). It also improves constructivist learning outcomes, with higher scores in prior knowledge activation (5.01 vs. 4.78), knowledge transfer (5.09 vs. 4.82), and error correction (5.24 vs. 4.68). We compare our improved RAG with other implementations and find that it outperforms the others with a 2.93% average improvement. In exercise recommendations, we achieve over 80% user satisfaction. As the number of recommended exercises increases, both SageJavon scores (

r = 0.68

) and final exam scores (

r = 0.62

) improve, demonstrating the positive impact of personalized recommendations. These results highlight SageJavon as a scalable, plug-and-play solution for programming education.

尽管编程教育取得了进步，但在管理认知负荷的同时提供适应性学习体验仍然是一个挑战。受Bloom分类法的启发，我们提出了学习-实践-评估-支持（LPES）框架，并将其集成到基于大型语言模型（llm）的人工智能辅导系统SageJavon中。SageJavon克服了传统AI导师的局限性，比如死板的内容传递和缺乏个性化的反馈。它包括四个创新：(1)用于个性化资源交付的增强型检索-增强生成（RAG）模型，(2)用于指导问题解决的启发式和后续问题，(3)自适应知识跟踪和推荐模型，以及(4)用于代码自动评估的FUPS-Score。在为期12周的课程中，有85名学生参与，SageJavon在知识问答模块中促进了5306次互动，在编程导师模块中促进了5144次互动，基于法学硕士的评分平均得分为87.39分和84.19分。与传统工具相比，SageJavon减少了认知负荷，在物理需求、时间需求和性能方面显示出改进的分数（所有p <； .001）。它也提高了建构主义学习成果，在先验知识激活方面得分更高(5.01 vs.；4.78)，知识转移(5.09 vs。4.82)和纠错(5.24 vs。4.68)。我们将改进后的RAG与其他实现进行了比较，发现它的平均改进率为2.93%，优于其他实现。在运动建议方面，我们达到了80%以上的用户满意度。随着推荐练习数量的增加，SageJavon分数（r=0.68）和期末考试分数（r=0.62）都有所提高，表明个性化推荐的积极影响。这些结果突出了sagejava作为一个可伸缩的、即插即用的编程教育解决方案。

{"title":"SageJavon: A scalable AI tutor for personalized programming learning","authors":"Hanyu Zhao , Yuzhuo Wu , Zhufeng Lu , Xiaohua Yu , Weikai Miao , Liangyu Chen","doi":"10.1016/j.ipm.2025.104605","DOIUrl":"10.1016/j.ipm.2025.104605","url":null,"abstract":"<div><div>Despite advancements in programming education, providing adaptive learning experiences while managing cognitive load remains a challenge. Inspired by Bloom’s Taxonomy, we propose the <strong>Learn-Practice-Evaluate-Support (LPES)</strong> framework, integrated into SageJavon, an AI-powered tutoring system based on Large Language Models (LLMs). SageJavon overcomes the limitations of traditional AI tutors, such as rigid content delivery and a lack of personalized feedback. It includes four innovations: (1) an enhanced Retrieval-Augmented Generation (RAG) model for personalized resource delivery, (2) heuristic and follow-up questions to guide problem-solving, (3) adaptive knowledge tracing and recommendation models, and (4) the FUPS-Score for automatic code assessment. Deployed in a 12-week course with 85 students, SageJavon facilitated 5306 interactions in the Knowledge Q&A module and 5144 in the Programming Mentor module, with LLM-based scoring yielding average scores of 87.39 and 84.19. Compared to traditional tools, SageJavon reduces cognitive load, showing improved scores in physical demand, temporal demand, and performance (all <em>p</em> < .001). It also improves constructivist learning outcomes, with higher scores in prior knowledge activation (5.01 vs. 4.78), knowledge transfer (5.09 vs. 4.82), and error correction (5.24 vs. 4.68). We compare our improved RAG with other implementations and find that it outperforms the others with a 2.93% average improvement. In exercise recommendations, we achieve over 80% user satisfaction. As the number of recommended exercises increases, both SageJavon scores (<span><math><mrow><mi>r</mi><mo>=</mo><mn>0.68</mn></mrow></math></span>) and final exam scores (<span><math><mrow><mi>r</mi><mo>=</mo><mn>0.62</mn></mrow></math></span>) improve, demonstrating the positive impact of personalized recommendations. These results highlight SageJavon as a scalable, plug-and-play solution for programming education.</div></div>","PeriodicalId":50365,"journal":{"name":"Information Processing & Management","volume":"63 5","pages":"Article 104605"},"PeriodicalIF":6.9,"publicationDate":"2026-02-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146190590","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"管理学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Defending LLMs against jailbreak attacks through representation offset detection 通过表示偏移检测保护llm免受越狱攻击

IF 6.9 1区管理学 Q1 COMPUTER SCIENCE, INFORMATION SYSTEMS

Information Processing & Management

Pub Date : 2026-02-02 DOI: 10.1016/j.ipm.2026.104662

Shuo Liu , Xiang Cheng , Zhenzhong Zheng , Sen Su

Jailbreak attacks bypass the security mechanisms of Large Language Models (LLMs) by disguising harmful prompts, seriously threatening model security. Existing approaches mainly rely on pre-training on specific datasets, which are usually costly and time-consuming. In this paper, we propose Representation Offset Defense (ROD), a plug-and-play detection framework that requires no pre-training. ROD identifies jailbreak attacks by exploiting the representational space mismatch between user inputs and their actual intents, and consists of two modules: main intent extraction (MIE) for generalizing the proposed schema, and representation offset analysis (ROA) for quantifying the semantic bias. We evaluate ROD with six jailbreak attack strategies on two widely used LLMs (Vicuna-7B and Llama2-7B). ROD achieves an average 96.1% defense success rate on Vicuna and 97.9% on Llama2, outperforming existing benchmarks.

越狱攻击绕过llm （Large Language Models）的安全机制，通过隐藏有害的提示信息，严重威胁llm的安全。现有的方法主要依赖于特定数据集的预训练，这通常是昂贵和耗时的。在本文中，我们提出了表征偏移防御（ROD），这是一种无需预训练的即插即用检测框架。ROD通过利用用户输入与其实际意图之间的表示空间不匹配来识别越狱攻击，它由两个模块组成：用于泛化提议模式的主意图提取（MIE）和用于量化语义偏差的表示偏移分析（ROA）。我们在两种广泛使用的llm （Vicuna-7B和Llama2-7B）上使用六种越狱攻击策略来评估ROD。ROD对骆马的平均防御成功率为96.1%，对羊驼的平均防御成功率为97.9%，优于现有的基准。

引用次数: 0

Feature structure co-optimized augmented network for graph anomaly detection 特征结构协同优化的增强网络图异常检测

IF 6.9 1区管理学 Q1 COMPUTER SCIENCE, INFORMATION SYSTEMS

Information Processing & Management

Pub Date : 2026-02-02 DOI: 10.1016/j.ipm.2026.104661

Yingyue Zhang, Huifang Ma, Rui Bing, Meihuizi Jia, Xiaofei Wang

Graph Anomaly Detection (GAD) aims to identify nodes that deviate from attribute or structural norms. Despite recent advances, redundant features and structural noise continue to obscure anomaly signals and bias node representations. We propose FSCAN (Feature-Structure Co-optimized Augmented Network), a plug-in module that jointly optimizes feature selection and graph refinement. FSCAN leverages spectral contribution-based selection, grounded in the Spiked Covariance Model and Singular Value Decomposition, to identify label-relevant features, while its adaptive refinement prunes noisy edges from head nodes and augments connections for tail nodes to enhance information propagation. Experiments on benchmark datasets show that FSCAN outperforms state-of-the-art baselines. Specifically, on the Questions dataset, when integrated with a supervised backbone, FSCAN achieves a GMean of 67.35%, representing a 6.63% improvement over the best baseline; when integrated with an unsupervised backbone, the GMean reaches 22.49%, corresponding to a 10.98% improvement over the baseline. Our code is publicly available at https://github.com/mala-tang/FSCAN.

图异常检测（GAD）旨在识别偏离属性或结构规范的节点。尽管最近取得了一些进展，但冗余特征和结构噪声仍然模糊了异常信号和偏置节点的表示。我们提出了FSCAN （feature - structure Co-optimized Augmented Network，特征结构协同优化增强网络），这是一个共同优化特征选择和图细化的插件模块。FSCAN利用基于谱贡献的选择，以尖峰协方差模型和奇异值分解为基础，识别标签相关特征，同时其自适应细化从头节点去除噪声边缘，并增加尾节点的连接，以增强信息传播。在基准数据集上的实验表明，FSCAN优于最先进的基线。具体来说，在Questions数据集上，当与监督主干集成时，FSCAN实现了67.35%的GMean，比最佳基线提高了6.63%；当与无监督骨干网集成时，GMean达到22.49%，比基线提高10.98%。我们的代码可以在https://github.com/mala-tang/FSCAN上公开获得。

{"title":"Feature structure co-optimized augmented network for graph anomaly detection","authors":"Yingyue Zhang, Huifang Ma, Rui Bing, Meihuizi Jia, Xiaofei Wang","doi":"10.1016/j.ipm.2026.104661","DOIUrl":"10.1016/j.ipm.2026.104661","url":null,"abstract":"<div><div>Graph Anomaly Detection (GAD) aims to identify nodes that deviate from attribute or structural norms. Despite recent advances, redundant features and structural noise continue to obscure anomaly signals and bias node representations. We propose FSCAN (Feature-Structure Co-optimized Augmented Network), a plug-in module that jointly optimizes feature selection and graph refinement. FSCAN leverages spectral contribution-based selection, grounded in the Spiked Covariance Model and Singular Value Decomposition, to identify label-relevant features, while its adaptive refinement prunes noisy edges from head nodes and augments connections for tail nodes to enhance information propagation. Experiments on benchmark datasets show that FSCAN outperforms state-of-the-art baselines. Specifically, on the Questions dataset, when integrated with a supervised backbone, FSCAN achieves a GMean of 67.35%, representing a 6.63% improvement over the best baseline; when integrated with an unsupervised backbone, the GMean reaches 22.49%, corresponding to a 10.98% improvement over the baseline. Our code is publicly available at <span><span>https://github.com/mala-tang/FSCAN</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50365,"journal":{"name":"Information Processing & Management","volume":"63 5","pages":"Article 104661"},"PeriodicalIF":6.9,"publicationDate":"2026-02-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146190594","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"管理学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Invariant learning improves out-of-distribution generalization for IP geolocation 不变学习改进了IP地理定位的分布外泛化

IF 6.9 1区管理学 Q1 COMPUTER SCIENCE, INFORMATION SYSTEMS

Information Processing & Management

Pub Date : 2026-01-29 DOI: 10.1016/j.ipm.2026.104641

Xiaohan Wang , Xueting Liu , Wenxin Tai , Joojo Walker , Yong Wang , Kai Chen , Fan Zhou

Accurate IP geolocation plays a critical role in a wide range of location-aware applications, from cybersecurity to content delivery. While recent advances in deep learning have led to substantial improvements in geolocation accuracy, existing methods often fail to generalize under out-of-distribution (OOD) scenarios caused by distribution shifts. To address this challenge, we propose a novel framework-Graph Invariant Learning (GIL)-for IP geolocation, referred to as GILGeo. Our approach is designed to identify invariant structural patterns in IP graphs across diverse environments, thereby enhancing model generalizability. By dynamically recombining invariant and spurious features, GILGeo simulates a variety of environmental conditions during training. This promotes the learning of domain-invariant representations and leads to significantly improved performance in unseen OOD settings. Extensive experiments on three real-world datasets show that GILGeo outperforms state-of-the-art baselines, establishing a new benchmark for IP geolocation under distributional shift. Our anonymized code and datasets are publicly available at: https://github.com/xiaohanwang01/GILGeo.

准确的IP地理定位在广泛的位置感知应用中起着至关重要的作用，从网络安全到内容交付。虽然深度学习的最新进展大大提高了地理定位的准确性，但现有的方法往往无法在分布变化引起的分布外（OOD）情况下进行泛化。为了解决这一挑战，我们提出了一个新的框架-图不变学习(GIL)-用于IP地理定位，称为GILGeo。我们的方法旨在识别跨不同环境的IP图中的不变结构模式，从而增强模型的可泛化性。通过动态重组不变和伪特征，GILGeo在训练过程中模拟各种环境条件。这促进了领域不变表示的学习，并显著提高了未见过的OOD设置的性能。在三个真实数据集上进行的大量实验表明，GILGeo优于最先进的基线，为分布移位下的IP地理定位建立了新的基准。我们的匿名代码和数据集公开可在：https://github.com/xiaohanwang01/GILGeo。

{"title":"Invariant learning improves out-of-distribution generalization for IP geolocation","authors":"Xiaohan Wang , Xueting Liu , Wenxin Tai , Joojo Walker , Yong Wang , Kai Chen , Fan Zhou","doi":"10.1016/j.ipm.2026.104641","DOIUrl":"10.1016/j.ipm.2026.104641","url":null,"abstract":"<div><div>Accurate IP geolocation plays a critical role in a wide range of location-aware applications, from cybersecurity to content delivery. While recent advances in deep learning have led to substantial improvements in geolocation accuracy, existing methods often fail to generalize under out-of-distribution (OOD) scenarios caused by distribution shifts. To address this challenge, we propose a novel framework-Graph Invariant Learning (GIL)-for IP geolocation, referred to as GILGeo. Our approach is designed to identify invariant structural patterns in IP graphs across diverse environments, thereby enhancing model generalizability. By dynamically recombining invariant and spurious features, GILGeo simulates a variety of environmental conditions during training. This promotes the learning of domain-invariant representations and leads to significantly improved performance in unseen OOD settings. Extensive experiments on three real-world datasets show that GILGeo outperforms state-of-the-art baselines, establishing a new benchmark for IP geolocation under distributional shift. Our anonymized code and datasets are publicly available at: <span><span>https://github.com/xiaohanwang01/GILGeo</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50365,"journal":{"name":"Information Processing & Management","volume":"63 5","pages":"Article 104641"},"PeriodicalIF":6.9,"publicationDate":"2026-01-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146081580","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"管理学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Modeling heterogeneous normality in time series anomaly detection 时间序列异常检测中的异构正态性建模

IF 6.9 1区管理学 Q1 COMPUTER SCIENCE, INFORMATION SYSTEMS

Information Processing & Management

Pub Date : 2026-01-28 DOI: 10.1016/j.ipm.2026.104644

Xiaohui Zhou , Yijie Wang , Hongzuo Xu , Yizhou Li

Time series anomaly detection is crucial in many fields, where the objective is to identify unusual patterns by learning normality from sequential observations. However, existing methods typically treat the entire training data as a single, homogeneous normal class, which disregards the normal diversity caused by distribution shifts over time. As a result, these methods are forced to learn a single, complex decision boundary that must enclose all variations of normal behavior, making it difficult to precisely distinguish subtle anomalies hidden within the normal patterns. Therefore, this paper tackles this challenge by explicitly modeling heterogeneous normality, which allows for learning simpler, localized decision boundaries to separate anomalies. Specifically, we propose a novel approach that decomposes the heterogeneous class space into multiple normal classes, adopting a two-stage coarse-to-fine training paradigm: (1) a Mixture of Experts (MoE) framework assigns pseudo-labels by routing input features to specialized experts for prediction, approximating the latent sub-class structure; (2) enhanced features are generated based on pseudo-labels and feature space is refined via spectral decomposition, which contracts class boundaries and better exposes anomalies. Extensive experiments on 23 univariate datasets and 17 multivariate datasets show that our approach significantly outperforms state-of-the-art competitors by 2.55%-21.76% in VUS-PR, validating the importance of modeling heterogeneous normality in time series anomaly detection.

时间序列异常检测在许多领域都是至关重要的，其目标是通过从序列观测中学习正态性来识别异常模式。然而，现有的方法通常将整个训练数据视为一个单一的、同构的正态类，而忽略了分布随时间变化而引起的正态多样性。因此，这些方法被迫学习一个单一的、复杂的决策边界，它必须包含正常行为的所有变化，这使得精确区分隐藏在正常模式中的微妙异常变得困难。因此，本文通过显式建模异构正态性来解决这一挑战，这允许学习更简单的局部决策边界来分离异常。具体来说，我们提出了一种新的方法，将异构类空间分解为多个正常类，采用两阶段粗到精的训练范式：(1)混合专家（MoE）框架通过将输入特征路由给专门的专家进行预测来分配伪标签，近似潜在的子类结构；(2)基于伪标签生成增强特征，通过谱分解细化特征空间，收缩类边界，更好地暴露异常；在23个单变量数据集和17个多变量数据集上进行的大量实验表明，我们的方法在VUS-PR方面明显优于最先进的竞争对手2.55%-21.76%，验证了异构正态性建模在时间序列异常检测中的重要性。

{"title":"Modeling heterogeneous normality in time series anomaly detection","authors":"Xiaohui Zhou , Yijie Wang , Hongzuo Xu , Yizhou Li","doi":"10.1016/j.ipm.2026.104644","DOIUrl":"10.1016/j.ipm.2026.104644","url":null,"abstract":"<div><div>Time series anomaly detection is crucial in many fields, where the objective is to identify unusual patterns by learning normality from sequential observations. However, existing methods typically treat the entire training data as a single, homogeneous normal class, which disregards the normal diversity caused by distribution shifts over time. As a result, these methods are forced to learn a single, complex decision boundary that must enclose all variations of normal behavior, making it difficult to precisely distinguish subtle anomalies hidden within the normal patterns. Therefore, this paper tackles this challenge by explicitly modeling heterogeneous normality, which allows for learning simpler, localized decision boundaries to separate anomalies. Specifically, we propose a novel approach that decomposes the heterogeneous class space into multiple normal classes, adopting a two-stage <em>coarse-to-fine</em> training paradigm: (1) a Mixture of Experts (MoE) framework assigns pseudo-labels by routing input features to specialized experts for prediction, approximating the latent sub-class structure; (2) enhanced features are generated based on pseudo-labels and feature space is refined via spectral decomposition, which contracts class boundaries and better exposes anomalies. Extensive experiments on 23 univariate datasets and 17 multivariate datasets show that our approach significantly outperforms state-of-the-art competitors by 2.55%-21.76% in VUS-PR, validating the importance of modeling heterogeneous normality in time series anomaly detection.</div></div>","PeriodicalId":50365,"journal":{"name":"Information Processing & Management","volume":"63 5","pages":"Article 104644"},"PeriodicalIF":6.9,"publicationDate":"2026-01-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146081583","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"管理学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Break fake frontiers: A triple-knowledge approach to multi-domain fake news detection 突破假新闻前沿：一种多领域假新闻检测的三重知识方法

IF 6.9 1区管理学 Q1 COMPUTER SCIENCE, INFORMATION SYSTEMS

Information Processing & Management

Pub Date : 2026-01-28 DOI: 10.1016/j.ipm.2026.104649

Xinnan Liu , Chenxi Liu , Anran Yu , Bin Tang , Zhenyang Cao , Zhengxiong Long , Runqi Su , Heng-Yang Lu

Nowdays, news events increasingly exhibit cross-domain topicality, making multi-domain fake news detection a critical yet challenging task. A key challenge is enhancing model performance for fake news detection while mitigating domain bias caused by unbalanced dataset. To address this, we propose the Knowledge-assistance Knowledge-mining Knowledge-debiased Multi-domain Fake News Detection Framework (K³MDFEND) which introduces external knowledge and proposes multi-domain contrastive learning. In particular, we integrate Large Language Models (LLMs) through a novel argumentation-based prompt engineering framework to obtain reliable external knowledge. We design Quality-Aware Attention Fusion module that dynamically weights evidence credibility to handle reviews of varying quality while combining with our deep learning framework. To further distill key insights from the comments while preserving their inherent semantic integrity, we leverage feature alignment techniques on the comment features. To further mitigate domain bias, we propose multi-domain contrastive learning and successfully combine the spurious correlations between domains and news authenticity. Extensive experiments on Chinese and English datasets demonstrate that K³MDFEND achieves state-of-the-art performance in both detection performance and bias metric reduction. On Chinese and English datasets, F1 scores, increase 92.89% ∼ 95.13% and 83.59% ∼ 85.28%, bias metric, decrease 0.8522 ∼ 0.5612 and 0.2698 ∼ 0.1931.

如今，新闻事件越来越呈现出跨领域的话题性，使得多领域假新闻检测成为一项关键而又具有挑战性的任务。一个关键的挑战是提高假新闻检测的模型性能，同时减轻由不平衡数据集引起的域偏差。为了解决这个问题，我们提出了知识辅助知识挖掘知识去偏见多域假新闻检测框架（k3mddefend），该框架引入了外部知识并提出了多域对比学习。特别是，我们通过一个新颖的基于论证的提示工程框架集成了大型语言模型（llm），以获得可靠的外部知识。我们设计了质量感知注意力融合模块，该模块动态加权证据可信度，以处理不同质量的评论，同时结合我们的深度学习框架。为了进一步从评论中提取关键的见解，同时保持其固有的语义完整性，我们在评论特性上利用了特征对齐技术。为了进一步减轻领域偏差，我们提出了多领域对比学习，并成功地将领域之间的虚假相关性与新闻真实性结合起来。在中文和英文数据集上的大量实验表明，k3mddefend在检测性能和减少偏置度量方面都达到了最先进的性能。在中文和英文数据集上，F1分数分别提高了92.89% ~ 95.13%和83.59% ~ 85.28%，偏差度量分别降低了0.8522 ~ 0.5612和0.2698 ~ 0.1931。

{"title":"Break fake frontiers: A triple-knowledge approach to multi-domain fake news detection","authors":"Xinnan Liu , Chenxi Liu , Anran Yu , Bin Tang , Zhenyang Cao , Zhengxiong Long , Runqi Su , Heng-Yang Lu","doi":"10.1016/j.ipm.2026.104649","DOIUrl":"10.1016/j.ipm.2026.104649","url":null,"abstract":"<div><div>Nowdays, news events increasingly exhibit cross-domain topicality, making multi-domain fake news detection a critical yet challenging task. A key challenge is enhancing model performance for fake news detection while mitigating domain bias caused by unbalanced dataset. To address this, we propose the Knowledge-assistance Knowledge-mining Knowledge-debiased Multi-domain Fake News Detection Framework (K<sup>3</sup>MDFEND) which introduces external knowledge and proposes multi-domain contrastive learning. In particular, we integrate Large Language Models (LLMs) through a novel argumentation-based prompt engineering framework to obtain reliable external knowledge. We design Quality-Aware Attention Fusion module that dynamically weights evidence credibility to handle reviews of varying quality while combining with our deep learning framework. To further distill key insights from the comments while preserving their inherent semantic integrity, we leverage feature alignment techniques on the comment features. To further mitigate domain bias, we propose multi-domain contrastive learning and successfully combine the spurious correlations between domains and news authenticity. Extensive experiments on Chinese and English datasets demonstrate that K<sup>3</sup>MDFEND achieves state-of-the-art performance in both detection performance and bias metric reduction. On Chinese and English datasets, F1 scores, increase 92.89% ∼ <strong>95.13%</strong> and 83.59% ∼ <strong>85.28%</strong>, bias metric, decrease 0.8522 ∼ <strong>0.5612</strong> and 0.2698 ∼ <strong>0.1931</strong>.</div></div>","PeriodicalId":50365,"journal":{"name":"Information Processing & Management","volume":"63 5","pages":"Article 104649"},"PeriodicalIF":6.9,"publicationDate":"2026-01-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146081579","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"管理学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0