Knowledge-Based Systems最新文献

{"title":"Frequency-spatial complementary attention network for computed tomography","authors":"Xing Wu ,&nbsp;Yimin Zhu ,&nbsp;Shuo Duan ,&nbsp;Xinyuan Zhang ,&nbsp;Xing Wei ,&nbsp;Bo Huang ,&nbsp;Quan Qian","doi":"10.1016/j.knosys.2026.115468","DOIUrl":"10.1016/j.knosys.2026.115468","url":null,"abstract":"<div><div>Computed tomography (CT) denoising is essential for clinical diagnosis and industrial inspection, but it is challenged by various noise and structural artifacts. Existing deep learning methods are limited by insufficient modeling of long-term dependencies, a disregard for intrinsic frequency-domain priors, and a significant domain gap caused by their reliance on unrealistic synthetic noise. To address these issues, a frequency-spatial complementary attention network (FSCANet) is proposed, which is based on the complementary fusion of frequency and spatial domains. The frequency domain branch explicitly decouples structural and phase information to model global context, while the spatial-domain branch improves local details. Simultaneously, a real-data-guided physics-informed noise model is introduced to bridge the domain gap by formalizing the physical noise generation process as a differentiable layer. FSCANet and the noise model are jointly optimized using a hybrid data-driven co-optimization strategy, resulting in a dynamic feedback loop that not only compels the noise model to generate physically interpretable noise but also drives FSCANet to achieve greater robustness. FSCANet achieves state-of-the-art performance on the DeepLesion dataset with a PSNR of 40.5861 dB and an SSIM of 0.9913, and demonstrates robust generalization on authentic clinical data from the Mayo dataset.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"338 ","pages":"Article 115468"},"PeriodicalIF":7.6,"publicationDate":"2026-04-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146174787","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"MM-AttacKG: A multimodal approach to attack graph construction with large language models","authors":"Yongheng Zhang ,&nbsp;Xinyun Zhao ,&nbsp;Yunshan Ma ,&nbsp;Haokai Ma ,&nbsp;Yingxiao Guan ,&nbsp;Guozheng Yang ,&nbsp;Yuliang Lu ,&nbsp;Xiang Wang","doi":"10.1016/j.knosys.2026.115483","DOIUrl":"10.1016/j.knosys.2026.115483","url":null,"abstract":"<div><div>Cyber Threat Intelligence (CTI) parsing aims to extract key threat information from massive data, transform it into actionable intelligence, enhance threat detection and defense efficiency, including attack graph construction, intelligence fusion, and indicator extraction. Among these research topics, Attack Graph Construction (AGC) is essential for visualizing and understanding the potential attack paths of threat events from CTI reports. Existing approaches primarily construct the attack graphs purely from the textual data to reveal the logical threat relationships between entities within the attack behavioral sequence. However, they typically overlook the specific threat information inherent in visual modalities, which preserves key threat details from inherently multimodal CTI reports. Inspired by the remarkable multimodal understanding capabilities of Multimodal Large Language Models (MLLMs), we explore their potential in enhancing multimodal attack graph construction. To be specific, we propose a novel framework, MM-AttacKG, which can effectively extract key information from threat images and integrate it into attack graph construction, thereby enhancing the comprehensiveness and accuracy of attack graphs. It first employs a threat image parsing module to extract critical threat information from images and generate textual descriptions using MLLMs. Subsequently, it builds an iterative question-answering pipeline tailored for image parsing to refine the understanding of threat images. Finally, it achieves content-level integration between attack graphs and image-based answers through MLLMs, completing threat information enhancement. We construct a new multimodal dataset, AG-LLM-mm, and conduct extensive experiments to evaluate the effectiveness of MM-AttacKG. The results demonstrate that MM-AttacKG can accurately identify key information in threat images and significantly improve the quality of multimodal attack graph construction, effectively addressing the shortcomings of existing methods in utilizing image-based threat information.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"338 ","pages":"Article 115483"},"PeriodicalIF":7.6,"publicationDate":"2026-04-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146175184","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Federated vision transformer with adaptive focal loss for medical image classification","authors":"Xinyuan Zhao ,&nbsp;Yihang Wu ,&nbsp;Ahmad Chaddad ,&nbsp;Tareef Daqqaq ,&nbsp;Reem Kateb","doi":"10.1016/j.knosys.2026.115474","DOIUrl":"10.1016/j.knosys.2026.115474","url":null,"abstract":"<div><div>While deep learning models like Vision Transformer (ViT) have achieved significant advances, they typically require large datasets. With data privacy regulations, access to many original datasets is restricted, especially medical images. Federated learning (FL) addresses this challenge by enabling global model aggregation without data exchange. However, the heterogeneity of the data and the class imbalance that exist in local clients pose challenges for the generalization of the model. This study proposes a FL framework leveraging a dynamic adaptive focal loss (DAFL) and a client-aware aggregation strategy for local training. Specifically, we design a dynamic class imbalance coefficient that adjusts based on each client’s sample distribution and class data distribution, ensuring minority classes receive sufficient attention and preventing sparse data from being ignored. To address client heterogeneity, a weighted aggregation strategy is adopted, which adapts to data size and characteristics to better capture inter-client variations. The classification results on three public datasets (ISIC, Ocular Disease and RSNA-ICH) show that the proposed framework outperforms DenseNet121, ResNet50, ViT-S/16, ViT-L/32, FedCLIP, Swin Transformer, CoAtNet, and MixNet in most cases, with accuracy improvements ranging from 0.98% to 41.69%. Ablation studies on the imbalanced ISIC dataset validate the effectiveness of the proposed loss function and aggregation strategy compared to traditional loss functions and other FL approaches. The codes can be found at: <span><span>https://github.com/AIPMLab/ViT-FLDAF</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"338 ","pages":"Article 115474"},"PeriodicalIF":7.6,"publicationDate":"2026-04-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146175230","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Multi-level dual contrastive learning for cloud API cold-start recommendation","authors":"Mengmeng Sun ,&nbsp;Yueshen Xu ,&nbsp;Dianlong You ,&nbsp;Zhen Chen","doi":"10.1016/j.knosys.2026.115534","DOIUrl":"10.1016/j.knosys.2026.115534","url":null,"abstract":"<div><div>A longstanding challenge in cloud API recommender systems is the cold-start problem associated with newly released cloud APIs, which lack historical interactions. Existing approaches typically 1) integrate the content and collaborative embeddings of a cloud API to generate its representation, or 2) adopt API-level alignment strategies that maximize mutual information between them. However, they often assume that cold cloud APIs have similar content features to warm ones, which doesn’t always hold in practice. Additionally, developers frequently combine multiple cloud APIs to create Mashups, indicating that focusing solely on individual cloud APIs fails to capture Mashup preferences. To this end, we propose a multi-level dual contrastive learning (MDCL) framework that explores Mashup preferences to impose effective embedding constraints on low-similarity cold cloud APIs. Specifically, MDCL generates a Mashup preference representation by aggregating the collaborative embeddings of warm cloud APIs based on the Mashup’s interaction history. It then performs group-level alignment between the content embedding of a cloud API and the Mashup’s preference representation, thereby guiding low-similarity cold cloud APIs toward the collaborative space. Furthermore, MDCL integrates API-level alignment and Mashup-API alignment to improve consistency between a cloud API’s content and collaborative embeddings, and to better model interaction patterns between Mashups and cloud APIs. A hybrid training strategy is employed to jointly optimize three alignment objectives: Mashup-API, API-level, and group-level alignment, achieving a better balance between cold-start and warm-start recommendations. Extensive experiments on real-world datasets demonstrate that MDCL outperforms SOTA methods in cold- and warm-start scenarios. Implementation code is available at <span><span>https://github.com/MengMeng3399/MDCL</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"338 ","pages":"Article 115534"},"PeriodicalIF":7.6,"publicationDate":"2026-04-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146175153","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Opposition and reinforcement learning growth-starfish optimization algorithm for engineering design and feature selection","authors":"Changting Zhong ,&nbsp;Hao Chen ,&nbsp;Dabo Xin ,&nbsp;Tong Xu ,&nbsp;Zeng Meng ,&nbsp;Xinwei Wang ,&nbsp;Ali Riza Yildiz ,&nbsp;Seyedali Mirjalili","doi":"10.1016/j.knosys.2026.115522","DOIUrl":"10.1016/j.knosys.2026.115522","url":null,"abstract":"<div><div>Starfish optimization algorithm (SFOA) is a bio-inspired metaheuristic algorithm for global optimization, which has demonstrated accuracy and efficiency in popular benchmark functions. However, for complex practical problems such as engineering design and feature selection, SFOA still requires a better balance between exploration and exploitation to ensure robust performance in real-world applications. In this paper, we present an improved SFOA algorithm named ORLGSFOA, which integrates opposition-based learning, reinforcement learning, and the growth optimizer with the basic SFOA. The algorithm first incorporates the opposition-based learning strategy during initialization to improve the diversity and quality of the initial solutions. Then, the updating rule from the growth optimizer is hybridized with SFOA to balance exploration and exploitation. Moreover, ORLGSFOA integrates the reinforcement learning strategy to reward the winner from SFOA and growth optimizer by adding updating positions during optimization to enhance global convergence. Experiments demonstrate the superior performance of ORLGSFOA. In comprehensive benchmark tests on 65 functions from classical, CEC2017, and CEC2022 suites, ORLGSFOA outperformed 15 other metaheuristic algorithms by achieving more accurate solutions. Additionally, this effectiveness translates directly to real-world applications, as is evidenced by tests on seven engineering design problems. Besides, the effectiveness of ORLGSFOA in solving discrete combinatorial optimization problems is verified through 52 feature selection problems, and the algorithm is extended to the wind engineering scenarios. In conclusion, ORLGSFOA demonstrates powerful efficacy in addressing a wide range of challenges, including global optimization, engineering design, and feature selection problems. The source code of ORLGSFOA is publicly available at: <span><span>https://ww2.mathworks.cn/matlabcentral/fileexchange/183223-orlgsfoa</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"338 ","pages":"Article 115522"},"PeriodicalIF":7.6,"publicationDate":"2026-04-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146175143","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Rethinking static weights: Language-guided adaptive weight adjustment for 3D visual grounding","authors":"Zongshun Wang ,&nbsp;Ce Li ,&nbsp;Zhiqiang Feng ,&nbsp;Limei Xiao ,&nbsp;Pengcheng Wang ,&nbsp;Mengmeng Ping","doi":"10.1016/j.knosys.2026.115467","DOIUrl":"10.1016/j.knosys.2026.115467","url":null,"abstract":"<div><div>3D Visual Grounding (3DVG) aims to accurately localize target objects in complex 3D point cloud scenes using natural language descriptions. However, current methods typically utilize static visual encoders with fixed parameters to handle the infinite variety of linguistic queries. This static approach inevitably leads to low signal-to-noise ratios in the feature inputs during the subsequent visual-language fusion stage. To overcome this limitation, we propose a Language-guided Adaptive Weight Adjustment (LAWA) framework that equips the visual backbone with query-aware dynamic adaptability during the early visual encoding stage via a lightweight language-guided strategy. Specifically, we first construct visual features that integrate class prior information using Object Semantic Augmented Encoding. Then, by leveraging weight coefficients derived from multimodal embeddings, we employ a Low-Rank Adaptation-based Dynamic Weight Adjustment (DWA) module to update the linear projection layers and weight matrices within the visual encoder’s attention mechanism. This approach enables the model to focus more effectively on visual regions that are semantically aligned with the textual descriptions. Extensive experiments demonstrate that LAWA achieves an [email protected] of 86.2% on the ScanRefer dataset, and overall accuracies of 69.5% and 58.4% on the Sr3D and Nr3D datasets, respectively, all while maintaining superior parameter efficiency.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"338 ","pages":"Article 115467"},"PeriodicalIF":7.6,"publicationDate":"2026-04-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146175187","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Analysis of double Beltrami horn surface resistor networks and efficient path planning","authors":"Xiaoyu Jiang ,&nbsp;Jianwei Dai ,&nbsp;Yanpeng Zheng ,&nbsp;Zhaolin Jiang","doi":"10.1016/j.knosys.2026.115489","DOIUrl":"10.1016/j.knosys.2026.115489","url":null,"abstract":"<div><div>Resistor networks, valued for their topological versatility and stable electrical properties, have emerged as a focal point across multiple disciplines. Yet, resistor networks with profound mathematical and physical significance remain largely unexplored. This study presents a detailed investigation of the double Beltrami horn surface resistor network and proposes an interpretable reasoning framework based on graph structures and grounded in physical laws. To improve the efficiency of large scale computation, the seventh type of discrete sine transform and Chebyshev polynomials of the first class are employed to derive the exact potential formula. In addition to generating potential distribution diagrams for various special scenarios, a fast algorithm is developed to significantly enhance the efficiency of potential computation. Furthermore, to expand the application potential of the resistor network, an efficient path planning algorithm based on the exact potential formula is proposed, and its applicability in dynamic environments is validated in preliminary experiments.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"338 ","pages":"Article 115489"},"PeriodicalIF":7.6,"publicationDate":"2026-04-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146175229","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"FeNeC: Enhancing continual learning via feature clustering with neighbor- or logit-based classification","authors":"Kamil Książek ,&nbsp;Hubert Jastrzębski ,&nbsp;Krzysztof Pniaczek ,&nbsp;Bartosz Trojan ,&nbsp;Michał Karp ,&nbsp;Jacek Tabor","doi":"10.1016/j.knosys.2026.115479","DOIUrl":"10.1016/j.knosys.2026.115479","url":null,"abstract":"<div><div>The ability of deep learning models to learn continuously is essential for adapting to new data categories and evolving data distributions. In recent years, approaches leveraging frozen feature extractors after an initial learning phase have been extensively studied. Many of these methods estimate per-class covariance matrices and prototypes based on backbone-derived feature representations. Within this paradigm, we introduce FeNeC (Feature Neighborhood Classifier) and FeNeC-Log, its variant based on the log-likelihood function. Our approach significantly extends the concept of per-class prototypes by constructing multiple, fine-grained sub-prototypes for each class, thereby enhancing the representation of class distributions. Utilizing the Mahalanobis distance, our models classify samples either through a nearest neighbor assignment to these sub-prototypes or trainable logit values assigned to consecutive classes. Our proposition can be seen as a generalization that reduces to existing single-prototype approaches in a special case, while extending them with the ability for more flexible adaptation to data. We demonstrate that our FeNeC variants establish state-of-the-art results across several benchmarks, proving particularly effective on CIFAR-100 and the complex ImageNet-Subset, where our method outperforms the strong FeCAM baseline by over 1% in average incremental accuracy and 1.5% in last task accuracy.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"338 ","pages":"Article 115479"},"PeriodicalIF":7.6,"publicationDate":"2026-04-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146174785","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A survey on anomaly segmentation in urban scene understanding with image data","authors":"Yuxuan Zhang ,&nbsp;Shuchang Wang ,&nbsp;Zhenbo Shi ,&nbsp;Wei Yang","doi":"10.1016/j.knosys.2026.115521","DOIUrl":"10.1016/j.knosys.2026.115521","url":null,"abstract":"<div><div>Semantic segmentation has made significant advancements over the past decade. However, it typically relies on a closed-set taxonomy, which limits its ability to generalize to objects of unknown categories. This limitation poses security risks in real-world applications, such as autonomous vehicles. To address this issue, anomaly segmentation in urban scene understanding has gained considerable attention, aiming to identify and segment outliers effectively. Considering the rapid progress in anomaly segmentation in recent years, there is no comprehensive survey of the latest developments in this field. In this paper, we systematically summarize recent advancements and introduce a novel perspective to categorize these approaches based on their underlying motivations. We then analyze the performance of each approach on several public leaderboards, demonstrating that this categorization criteria reflects the development trends of recent progress. Additionally, we identify existing challenges and outlook for potential future research directions.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"338 ","pages":"Article 115521"},"PeriodicalIF":7.6,"publicationDate":"2026-04-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146175154","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Boundary-aware and multi-angle modeling-based object tracking in polarimetric images","authors":"Qiaohui Wang ,&nbsp;Fan Shi ,&nbsp;Mianzhao Wang ,&nbsp;Xinbo Geng ,&nbsp;Meng Zhao","doi":"10.1016/j.knosys.2026.115442","DOIUrl":"10.1016/j.knosys.2026.115442","url":null,"abstract":"<div><div>Object tracking is a fundamental task in computer vision with applications ranging from surveillance to autonomous driving. Although RGB-based tracking methods have seen significant advancements by leveraging color and texture features, they often struggle under challenging conditions such as low light, occlusions, and fast motion. Polarimetric imaging, which encodes surface properties, material characteristics, and geometric structures, offers unique advantages as a complementary modality. However, its potential remains underexplored due to the lack of large-scale datasets and specialized algorithms designed for polarization-specific features. To address this gap, we introduce POL, the first large-scale benchmark dataset for polarimetric vision that enables comprehensive evaluations under diverse conditions. Building on this dataset, we propose PMTT, a cross-modal transformer framework that integrates polarimetric and RGB data. The Detailed Feature Prompter (DFP) module extracts boundary and multi-angle features from polarimetric images, while the Spatial-Channel Attention (SCA) mechanism enhances feature recognition in complex environments. Extensive experiments confirm that PMTT superior performance and robustness, highlighting the transformative potential of polarimetric imaging for dynamic object tracking.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"338 ","pages":"Article 115442"},"PeriodicalIF":7.6,"publicationDate":"2026-04-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146175321","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}