Pub Date : 2026-01-23DOI: 10.1016/j.ins.2026.123143
Nikolay I. Kalmykov , Razan Dibo , Kaiyu Shen , Zhonghan Xu , Anh-Huy Phan , Yipeng Liu , Ivan Oseledets
Neural image compression (NIC) has become the state-of-the-art for rate-distortion performance, yet its security vulnerabilities remain significantly less understood than those of classifiers. Existing adversarial attacks on NICs are often naive adaptations of pixel-space methods, overlooking the unique, structured nature of the compression pipeline. In this work, we propose a more advanced class of vulnerabilities by introducing T-MLA, the first targeted multiscale log–exponential attack framework. We introduce adversarial perturbations in the wavelet domain that concentrate on less perceptually salient coefficients, improving the stealth of the attack. Extensive evaluation across multiple state-of-the-art NIC architectures on standard image compression benchmarks reveals a large drop in reconstruction quality while the perturbations remain visually imperceptible. On standard NIC benchmarks, T-MLA achieves targeted degradation of reconstruction quality while improving perturbation imperceptibility (higher PSNR/VIF of the perturbed inputs) compared to PGD-style baselines at comparable attack success, as summarized in our main results. Our findings reveal a critical security flaw at the core of generative and content delivery pipelines.
{"title":"T-MLA: A targeted multiscale log–exponential attack framework for neural image compression","authors":"Nikolay I. Kalmykov , Razan Dibo , Kaiyu Shen , Zhonghan Xu , Anh-Huy Phan , Yipeng Liu , Ivan Oseledets","doi":"10.1016/j.ins.2026.123143","DOIUrl":"10.1016/j.ins.2026.123143","url":null,"abstract":"<div><div>Neural image compression (NIC) has become the state-of-the-art for rate-distortion performance, yet its security vulnerabilities remain significantly less understood than those of classifiers. Existing adversarial attacks on NICs are often naive adaptations of pixel-space methods, overlooking the unique, structured nature of the compression pipeline. In this work, we propose a more advanced class of vulnerabilities by introducing T-MLA, the first targeted multiscale log–exponential attack framework. We introduce adversarial perturbations in the wavelet domain that concentrate on less perceptually salient coefficients, improving the stealth of the attack. Extensive evaluation across multiple state-of-the-art NIC architectures on standard image compression benchmarks reveals a large drop in reconstruction quality while the perturbations remain visually imperceptible. On standard NIC benchmarks, T-MLA achieves targeted degradation of reconstruction quality while improving perturbation imperceptibility (higher PSNR/VIF of the perturbed inputs) compared to PGD-style baselines at comparable attack success, as summarized in our main results. Our findings reveal a critical security flaw at the core of generative and content delivery pipelines.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"738 ","pages":"Article 123143"},"PeriodicalIF":6.8,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146081295","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}
Pub Date : 2026-01-23DOI: 10.1016/j.ins.2026.123131
Aijun Yan , Xin Liu
In the modeling process of model predictive control (MPC), the model typically exhibits non-convex characteristics, which make the optimization problem complex and prone to local optima. To address this, an MPC modeling method based on input convex stochastic configuration networks (SCN) is proposed. The method imposes convexity constraints on both network architecture and activation functions. A Sparsemax-based activation function selection mechanism is developed to adaptively choose convex activation functions for each configuration node. Output weights are determined using the alternating direction method of multipliers to solve least-squares problems with non-negative constraints. Two architectures are constructed: fully input convex and partially input convex SCN. Through a dynamic supervision mechanism, it is theoretically proven that the proposed model approximates convex functions to arbitrary accuracy under weight constraints. Experimental results demonstrate improved fitting accuracy with convex approximation guarantees, and control examples show enhanced closed-loop performance by ensuring MPC optimization convexity.
{"title":"Input convex stochastic configuration networks modeling method for predictive control","authors":"Aijun Yan , Xin Liu","doi":"10.1016/j.ins.2026.123131","DOIUrl":"10.1016/j.ins.2026.123131","url":null,"abstract":"<div><div>In the modeling process of model predictive control (MPC), the model typically exhibits non-convex characteristics, which make the optimization problem complex and prone to local optima. To address this, an MPC modeling method based on input convex stochastic configuration networks (SCN) is proposed. The method imposes convexity constraints on both network architecture and activation functions. A Sparsemax-based activation function selection mechanism is developed to adaptively choose convex activation functions for each configuration node. Output weights are determined using the alternating direction method of multipliers to solve least-squares problems with non-negative constraints. Two architectures are constructed: fully input convex and partially input convex SCN. Through a dynamic supervision mechanism, it is theoretically proven that the proposed model approximates convex functions to arbitrary accuracy under weight constraints. Experimental results demonstrate improved fitting accuracy with convex approximation guarantees, and control examples show enhanced closed-loop performance by ensuring MPC optimization convexity.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"738 ","pages":"Article 123131"},"PeriodicalIF":6.8,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146081297","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}
Pub Date : 2026-01-23DOI: 10.1016/j.ins.2026.123141
Rongna Cai , Haibin Ouyang , Steven Li , Gaige Wang , Weiping Ding
To tackle challenges in industrial image defect detection, guided by three core hypotheses: dataset representativeness, continuous differentiable NAS search space, and GPU-based computing environment, this study presents an enhanced particle swarm optimization (PSO)-based neural architecture search (NAS) method designated as DNE-PSO-NAS. Firstly, it employs a two-level binary particle encoding scheme for network layer configurations and connectivity, transforming architecture search into a multi-dimensional optimization problem. Secondly, an improved MBConv module with CBAM is developed to reinforce the model’s ability to perceive local and global features of defects, thereby raising the signal-to-noise ratio for tiny defect regions. Additionally, dynamic ring neighborhood velocity topology and swarm entropy-driven mutation are proposed to balance exploration and exploitation, boosting PSO’s optimization efficiency. Finally, a low-fidelity evaluation strategy is incorporated, forming a three-stage framework that reduces input space via image downsampling, compresses convolutional layer parameters to lower spatial complexity, and adopts a dynamic training termination mechanism based on fitness tracking. Experiments on NEU-DET and WM-811 K datasets demonstrate that its discovered architectures surpass traditional CNNs and SOTA methods, with classification accuracy reaching 100% on NEU-DET and 93% on WM-811K. Meanwhile, our algorithm cuts computational costs significantly and the results highlight major benefits for real-time industrial quality inspection.
{"title":"Neural architecture search using an enhanced particle swarm optimization algorithm for industrial image classification","authors":"Rongna Cai , Haibin Ouyang , Steven Li , Gaige Wang , Weiping Ding","doi":"10.1016/j.ins.2026.123141","DOIUrl":"10.1016/j.ins.2026.123141","url":null,"abstract":"<div><div>To tackle challenges in industrial image defect detection, guided by three core hypotheses: dataset representativeness, continuous differentiable NAS search space, and GPU-based computing environment, this study presents an enhanced particle swarm optimization (PSO)-based neural architecture search (NAS) method designated as DNE-PSO-NAS. Firstly, it employs a two-level binary particle encoding scheme for network layer configurations and connectivity, transforming architecture search into a multi-dimensional optimization problem. Secondly, an improved MBConv module with CBAM is developed to reinforce the model’s ability to perceive local and global features of defects, thereby raising the signal-to-noise ratio for tiny defect regions. Additionally, dynamic ring neighborhood velocity topology and swarm entropy-driven mutation are proposed to balance exploration and exploitation, boosting PSO’s optimization efficiency. Finally, a low-fidelity evaluation strategy is incorporated, forming a three-stage framework that reduces input space via image downsampling, compresses convolutional layer parameters to lower spatial complexity, and adopts a dynamic training termination mechanism based on fitness tracking. Experiments on NEU-DET and WM-811 K datasets demonstrate that its discovered architectures surpass traditional CNNs and SOTA methods, with classification accuracy reaching 100% on NEU-DET and 93% on WM-811K. Meanwhile, our algorithm cuts computational costs significantly and the results highlight major benefits for real-time industrial quality inspection.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"738 ","pages":"Article 123141"},"PeriodicalIF":6.8,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146043324","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}
Pub Date : 2026-01-22DOI: 10.1016/j.ins.2026.123128
El Arbi Abdellaoui Alaoui , Hayat Sahlaoui , Amine Sallah , Anand Nayyar
The intractable nature of the Kernel SHAP computation presents a significant barrier to implementing rigorous explainable AI in practical large scale machine learning systems. The standard SHAP calculations are based on iterative evaluations of the model averaged over all possible feature subsets, and they become intractable for large dimensional data. To mitigate this difficulty, this paper presents “DE-SHAP”, an evolutionary optimization approach which speeds up ES-method by automatically tuning its parameters via Differential Evolution (DE). DE systematically optimizes two critical parameters background dataset size and Monte Carlo sampling count to minimize computational cost and maintain theoretical soundness within SHAP’s additive feature attribution framework. The proposed framework employs specialized evolutionary operators to ensure convergence efficiency and stability during optimization.
To test and validate the proposed methodology, extensive experiments were performed on various benchmark datasets and model architectures, and the results showed that DE-SHAP reduces computing cost by 52–97%, while the deviation of SHAP values is less than 5% and accuracy of the model remains within a range of approximately 0.5%. Given the fact that it is usually expensive to obtain explanations compared to predictions, these findings provide evidence that DE-SHAP can offer a similar quality of interpretability for a much lower computational cost. By implementing a computationally efficient theoretically justified improvement to a popular interpretability approach. DE-SHAP enables scalable and practical deployment of high-quality model explanations for complex systems with up to 784 input features. This contribution advances the feasibility of rigorous explainable AI in real-world applications, bridging the gap between interpretability research and operational machine learning.
{"title":"DE-SHAP: A meta-optimization framework leveraging differential evolution for efficient and scalable kernel SHAP explanations","authors":"El Arbi Abdellaoui Alaoui , Hayat Sahlaoui , Amine Sallah , Anand Nayyar","doi":"10.1016/j.ins.2026.123128","DOIUrl":"10.1016/j.ins.2026.123128","url":null,"abstract":"<div><div>The intractable nature of the Kernel SHAP computation presents a significant barrier to implementing rigorous explainable AI in practical large scale machine learning systems. The standard SHAP calculations are based on iterative evaluations of the model averaged over all possible feature subsets, and they become intractable for large dimensional data. To mitigate this difficulty, this paper presents “DE-SHAP”, an evolutionary optimization approach which speeds up ES-method by automatically tuning its parameters via Differential Evolution (DE). DE systematically optimizes two critical parameters background dataset size and Monte Carlo sampling count to minimize computational cost and maintain theoretical soundness within SHAP’s additive feature attribution framework. The proposed framework employs specialized evolutionary operators to ensure convergence efficiency and stability during optimization.</div><div>To test and validate the proposed methodology, extensive experiments were performed on various benchmark datasets and model architectures, and the results showed that DE-SHAP reduces computing cost by 52–97%, while the deviation of SHAP values is less than 5% and accuracy of the model remains within a range of approximately 0.5%. Given the fact that it is usually expensive to obtain explanations compared to predictions, these findings provide evidence that DE-SHAP can offer a similar quality of interpretability for a much lower computational cost. By implementing a computationally efficient theoretically justified improvement to a popular interpretability approach. DE-SHAP enables scalable and practical deployment of high-quality model explanations for complex systems with up to 784 input features. This contribution advances the feasibility of rigorous explainable AI in real-world applications, bridging the gap between interpretability research and operational machine learning.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"739 ","pages":"Article 123128"},"PeriodicalIF":6.8,"publicationDate":"2026-01-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146190946","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}
Pub Date : 2026-01-22DOI: 10.1016/j.ins.2026.123134
Haipeng Cui , Kai Xiao , Hua Wang , Xuxin Zhang
High-precision trajectory prediction can promote safe and efficient autonomous driving decisions. Existing state-of-the-art models, such as Dual-Attention Mechanism (DAM) and Hierarchical Attention Network (HAN), treat all neighboring vehicles as undifferentiated sets, ignoring lane structures when extracting spatial features. In this study, we propose a novel Lane-specific Spatial-Temporal Attention Network (LSTAN) to address the lane-level traffic information in vehicle trajectory prediction. Specifically, we employ an encoder module based on a Long Short-Term Memory Network to extract temporal features for target vehicles and their surrounding vehicles. Meanwhile, a lane attention module (LAM) and a temporal self-attention module (TSAM) are proposed for spatial and temporal feature extractions. The LAM introduces a dual-attention framework to discern spatial relationships between the target vehicle and its surrounding vehicles considering the lane-level effects. The TSAM refines the temporal features for target vehicles. Finally, the decoder integrates the learned features with the driving intention to obtain the predicted trajectories. Experiments are conducted using two real-world datasets: the next generation simulation (NGSIM) and HighD. Results show that the LSTAN outperforms the benchmarks by an average root mean square error (RMSE) of 0.37 m. Ablation studies and component replacement experiments are conducted to evaluate the effectiveness of the components in LSTAN.
{"title":"Lane-flow-learning based autonomous vehicle trajectory prediction using spatial–temporal fusion attention","authors":"Haipeng Cui , Kai Xiao , Hua Wang , Xuxin Zhang","doi":"10.1016/j.ins.2026.123134","DOIUrl":"10.1016/j.ins.2026.123134","url":null,"abstract":"<div><div>High-precision trajectory prediction can promote safe and efficient autonomous driving decisions. Existing state-of-the-art models, such as Dual-Attention Mechanism (DAM) and Hierarchical Attention Network (HAN), treat all neighboring vehicles as undifferentiated sets, ignoring lane structures when extracting spatial features. In this study, we propose a novel Lane-specific Spatial-Temporal Attention Network (LSTAN) to address the lane-level traffic information in vehicle trajectory prediction. Specifically, we employ an encoder module based on a Long Short-Term Memory Network to extract temporal features for target vehicles and their surrounding vehicles. Meanwhile, a lane attention module (LAM) and a temporal self-attention module (TSAM) are proposed for spatial and temporal feature extractions. The LAM introduces a dual-attention framework to discern spatial relationships between the target vehicle and its surrounding vehicles considering the lane-level effects. The TSAM refines the temporal features for target vehicles. Finally, the decoder integrates the learned features with the driving intention to obtain the predicted trajectories. Experiments are conducted using two real-world datasets: the next generation simulation (NGSIM) and HighD. Results show that the LSTAN outperforms the benchmarks by an average root mean square error (RMSE) of 0.37 m. Ablation studies and component replacement experiments are conducted to evaluate the effectiveness of the components in LSTAN.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"737 ","pages":"Article 123134"},"PeriodicalIF":6.8,"publicationDate":"2026-01-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080724","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}
Pub Date : 2026-01-22DOI: 10.1016/j.ins.2026.123133
Muhammad Imran Khalid , Jian-Xun Mi , Ghulam Ali , Tariq Ali , Mohammad Hijji , Muhammad Ayaz , Zia-ur-Rehman
The complexity and opaque internal mechanisms of deep learning models, particularly modern object detectors like DETR, make them challenging to interpret. Existing explainability methods, such as ODAM, produce spatial heatmaps but often fail to distinguish overlapping objects or convey semantic meaning. Concept-based methods, while interpretable, typically lack precise instance-level localization. To overcome these limitations, we propose IntegraXAI (Integrated Explainable AI), a novel framework that, for the first time, integrates two distinct XAI modalities for object detection: (1) gradient-based, instance-specific heatmaps (inspired by ODAM) for spatial localization, and (2) semantic concept discovery via Non-negative Matrix Factorization (NMF) with concept importance quantification using Sobol sensitivity analysis. The proposed three-stage framework provides insight not only into where the detector focuses its attention, but also into the semantic cues that ultimately guide its predictions. The effectiveness of IntegraXAI is validated across multiple object detection architectures, including DETR, YOLOv5, and Faster R-CNN, using the COCO benchmark dataset. Experimental findings show that the proposed method consistently outperforms existing explainability techniques, including Grad-CAM++, D-RISE, and individual ODAM or CRAFT variants, achieving higher spatial localization accuracy and clearer semantic interpretation. At the same time, IntegraXAI maintains stable, practical computational requirements, producing explanations in approximately 1 s per image, which is substantially more efficient than perturbation-based approaches like D-RISE. By jointly integrating spatial, semantic, and quantitative explanation mechanisms, the proposed framework improves the interpretability and trustworthiness of object detection systems, particularly in safety–critical domains such as autonomous driving and video surveillance.
{"title":"Interpretable object detection via integrated heatmap, concept attribution, and sobol sensitivity analysis","authors":"Muhammad Imran Khalid , Jian-Xun Mi , Ghulam Ali , Tariq Ali , Mohammad Hijji , Muhammad Ayaz , Zia-ur-Rehman","doi":"10.1016/j.ins.2026.123133","DOIUrl":"10.1016/j.ins.2026.123133","url":null,"abstract":"<div><div>The complexity and opaque internal mechanisms of deep learning models, particularly modern object detectors like DETR, make them challenging to interpret. Existing explainability methods, such as ODAM, produce spatial heatmaps but often fail to distinguish overlapping objects or convey semantic meaning. Concept-based methods, while interpretable, typically lack precise instance-level localization. To overcome these limitations, we propose IntegraXAI (Integrated Explainable AI), a novel framework that, for the first time, integrates two distinct XAI modalities for object detection: (1) gradient-based, instance-specific heatmaps (inspired by ODAM) for spatial localization, and (2) semantic concept discovery via Non-negative Matrix Factorization (NMF) with concept importance quantification using Sobol sensitivity analysis. The proposed three-stage framework provides insight not only into where the detector focuses its attention, but also into the semantic cues that ultimately guide its predictions. The effectiveness of IntegraXAI is validated across multiple object detection architectures, including DETR, YOLOv5, and Faster R-CNN, using the COCO benchmark dataset. Experimental findings show that the proposed method consistently outperforms existing explainability techniques, including Grad-CAM++, D-RISE, and individual ODAM or CRAFT variants, achieving higher spatial localization accuracy and clearer semantic interpretation. At the same time, IntegraXAI maintains stable, practical computational requirements, producing explanations in approximately 1 s per image, which is substantially more efficient than perturbation-based approaches like D-RISE. By jointly integrating spatial, semantic, and quantitative explanation mechanisms, the proposed framework improves the interpretability and trustworthiness of object detection systems, particularly in safety–critical domains such as autonomous driving and video surveillance.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"738 ","pages":"Article 123133"},"PeriodicalIF":6.8,"publicationDate":"2026-01-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146081296","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}
Pub Date : 2026-01-22DOI: 10.1016/j.ins.2026.123130
Jun Wan , Xinyu Xiong , Ning Chen , Zhihui Lai , Jie Zhou , Wenwen Min
Recently, deep learning based facial landmark detection (FLD) methods have achieved considerable success. However, in challenging scenarios such as large pose variations, illumination changes, and facial expression variations, they still struggle to accurately capture the geometric structure of the face, resulting in performance degradation. Moreover, the limited size and diversity of existing FLD datasets hinder robust model training, leading to reduced detection accuracy. To address these challenges, we propose a Frequency-Guided Task-Balancing Transformer (FGTBT), which enhances facial structure perception through frequency-domain modeling and multi-dataset unified training. Specifically, we propose a novel Fine-Grained Multi-Task Balancing loss (FMB-loss), which moves beyond coarse task-level balancing by assigning weights to individual landmarks based on their occurrence across datasets. This enables more effective unified training and mitigates the issue of inconsistent gradient magnitudes. Additionally, a Frequency-Guided Structure-Aware (FGSA) model is designed to utilize frequency-guided structure injection and regularization to help learn facial structure constraints. Extensive experimental results on popular benchmark datasets demonstrate that the integration of the proposed FMB-loss and FGSA model into our FGTBT framework achieves performance comparable to state-of-the-art methods. The code is available at https://github.com/Xi0ngxinyu/FGTBT.
{"title":"FGTBT: Frequency-guided task-balancing transformer for unified facial landmark detection","authors":"Jun Wan , Xinyu Xiong , Ning Chen , Zhihui Lai , Jie Zhou , Wenwen Min","doi":"10.1016/j.ins.2026.123130","DOIUrl":"10.1016/j.ins.2026.123130","url":null,"abstract":"<div><div>Recently, deep learning based facial landmark detection (FLD) methods have achieved considerable success. However, in challenging scenarios such as large pose variations, illumination changes, and facial expression variations, they still struggle to accurately capture the geometric structure of the face, resulting in performance degradation. Moreover, the limited size and diversity of existing FLD datasets hinder robust model training, leading to reduced detection accuracy. To address these challenges, we propose a Frequency-Guided Task-Balancing Transformer (FGTBT), which enhances facial structure perception through frequency-domain modeling and multi-dataset unified training. Specifically, we propose a novel Fine-Grained Multi-Task Balancing loss (FMB-loss), which moves beyond coarse task-level balancing by assigning weights to individual landmarks based on their occurrence across datasets. This enables more effective unified training and mitigates the issue of inconsistent gradient magnitudes. Additionally, a Frequency-Guided Structure-Aware (FGSA) model is designed to utilize frequency-guided structure injection and regularization to help learn facial structure constraints. Extensive experimental results on popular benchmark datasets demonstrate that the integration of the proposed FMB-loss and FGSA model into our FGTBT framework achieves performance comparable to state-of-the-art methods. The code is available at <span><span>https://github.com/Xi0ngxinyu/FGTBT</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"738 ","pages":"Article 123130"},"PeriodicalIF":6.8,"publicationDate":"2026-01-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146081299","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}
Pub Date : 2026-01-21DOI: 10.1016/j.ins.2026.123140
Francesco Villani , Dario Lazzaro , Antonio Emanuele Cinà , Matteo Dell’Amico , Battista Biggio , Fabio Roli
Data poisoning attacks on clustering algorithms have received limited attention, with existing methods struggling to scale efficiently as dataset sizes and feature counts increase. These attacks typically require re-clustering the entire dataset multiple times to generate predictions and assess the attacker’s objectives, significantly hindering their scalability. This paper addresses these limitations by proposing Sonic, a novel genetic data poisoning attack that leverages incremental and scalable clustering algorithms, e.g., FISHDBC, as surrogates to accelerate poisoning attacks against graph-based and density-based clustering methods, such as HDBSCAN. We empirically demonstrate the effectiveness and efficiency of Sonic in poisoning the target clustering algorithms. We then conduct a comprehensive analysis of the factors affecting the scalability and transferability of poisoning attacks against clustering algorithms, and we conclude by examining the robustness of hyperparameters in our attack strategy Sonic.
{"title":"Sonic: Fast and transferable data poisoning on clustering algorithms","authors":"Francesco Villani , Dario Lazzaro , Antonio Emanuele Cinà , Matteo Dell’Amico , Battista Biggio , Fabio Roli","doi":"10.1016/j.ins.2026.123140","DOIUrl":"10.1016/j.ins.2026.123140","url":null,"abstract":"<div><div>Data poisoning attacks on clustering algorithms have received limited attention, with existing methods struggling to scale efficiently as dataset sizes and feature counts increase. These attacks typically require re-clustering the entire dataset multiple times to generate predictions and assess the attacker’s objectives, significantly hindering their scalability. This paper addresses these limitations by proposing <span>Sonic</span>, a novel genetic data poisoning attack that leverages incremental and scalable clustering algorithms, e.g., FISHDBC, as surrogates to accelerate poisoning attacks against graph-based and density-based clustering methods, such as HDBSCAN. We empirically demonstrate the effectiveness and efficiency of <span>Sonic</span> in poisoning the target clustering algorithms. We then conduct a comprehensive analysis of the factors affecting the scalability and transferability of poisoning attacks against clustering algorithms, and we conclude by examining the robustness of hyperparameters in our attack strategy <span>Sonic</span>.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"738 ","pages":"Article 123140"},"PeriodicalIF":6.8,"publicationDate":"2026-01-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146081293","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}
Pub Date : 2026-01-21DOI: 10.1016/j.ins.2026.123138
Ziqi Meng , Wentao Fan , Bo Wang , Chunlin Chen , Huaxiong Li
The popular embedded feature selection approaches generally incorporate feature selection into a classification or regression task with sparse learning. In data mining, feature selection serves as an essential process. Due to the common scarcity of label information, unsupervised feature selection has attracted increasing attention. Most current methods face two challenges. Firstly, a vast majority of them rely on discovering the similarity relationships among samples to guide feature selection, which limits their efficiency and scalability due to the high time consumption of similarity graph learning. Secondly, they generally explore the data in the original or a fixed low-dimensional space, i.e., from a single-view perspective, which may not sufficiently exploit the underlying information. To address these issues, a novel diverse Embeddings and consensus Pseudo-supervision based unsupervised Feature Selection method, i.e., EPFS, is proposed in this paper, which solves the problem from a multi-view perspective in an efficient way. The EPFS framework integrates latent embedding learning, consensus pseudo-label learning, and sparse feature selection, enabling their mutual reinforcement and synergistic enhancement. For enhancing the pseudo-label quality, EPFS generates multiple distinct latent embeddings by mapping the original data into heterogeneous informative subspaces with simultaneous encoder–decoder reconstruction loss minimization. An auto-weighted collaboration strategy is adopted to learn a consensus pseudo-label matrix by using diverse embeddings. The sparse feature selection process is seamlessly incorporated into the framework. With an efficient linear-time algorithm, our model surpasses existing state-of-the-art approaches in experimental evaluations.
{"title":"Diverse embeddings and consensus pseudo-supervision learning for unsupervised feature selection","authors":"Ziqi Meng , Wentao Fan , Bo Wang , Chunlin Chen , Huaxiong Li","doi":"10.1016/j.ins.2026.123138","DOIUrl":"10.1016/j.ins.2026.123138","url":null,"abstract":"<div><div>The popular embedded feature selection approaches generally incorporate feature selection into a classification or regression task with sparse learning. In data mining, feature selection serves as an essential process. Due to the common scarcity of label information, unsupervised feature selection has attracted increasing attention. Most current methods face two challenges. Firstly, a vast majority of them rely on discovering the similarity relationships among samples to guide feature selection, which limits their efficiency and scalability due to the high time consumption of similarity graph learning. Secondly, they generally explore the data in the original or a fixed low-dimensional space, i.e., from a single-view perspective, which may not sufficiently exploit the underlying information. To address these issues, a novel diverse Embeddings and consensus Pseudo-supervision based unsupervised Feature Selection method, i.e., EPFS, is proposed in this paper, which solves the problem from a multi-view perspective in an efficient way. The EPFS framework integrates latent embedding learning, consensus pseudo-label learning, and sparse feature selection, enabling their mutual reinforcement and synergistic enhancement. For enhancing the pseudo-label quality, EPFS generates multiple distinct latent embeddings by mapping the original data into heterogeneous informative subspaces with simultaneous encoder–decoder reconstruction loss minimization. An auto-weighted collaboration strategy is adopted to learn a consensus pseudo-label matrix by using diverse embeddings. The sparse feature selection process is seamlessly incorporated into the framework. With an efficient linear-time algorithm, our model surpasses existing state-of-the-art approaches in experimental evaluations.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"739 ","pages":"Article 123138"},"PeriodicalIF":6.8,"publicationDate":"2026-01-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146081917","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}
Pub Date : 2026-01-21DOI: 10.1016/j.ins.2026.123127
Wenli Chen , Xiaojian Li
This study investigates the problem of data-driven fault-tolerant output synchronization control for heterogeneous multi-agent systems with unknown dynamics. Unlike the existing approaches that rely on prior knowledge of system matrices, this work proposes a novel method to design distributed data-driven fault-tolerant output synchronization controllers using data and output regulation theory. The output regulator equations are essential for fault-tolerant output synchronization control, whereas exact solutions from noise-corrupted data are challenging to obtain. To address this issue, a data-driven optimization problem is formulated to seek approximate solutions by minimizing the output regulation error matrices. Stability conditions are then derived in the form of data-dependent programs, whose solutions directly yield stabilizing feedback gains for agents. This approach ensures the achievement of output synchronization by utilizing data. Furthermore, a data-driven fault-tolerant controller is constructed by integrating adaptive control techniques with approximate solutions to the output regulator equations and stabilizing feedback gains learned from data, equipping agents with fault-tolerant capabilities. Theoretical analysis demonstrates that the proposed controller ensures the output synchronization errors are globally ultimately bounded (GUB). To validate the theoretical results, simulation examples are presented to demonstrate their efficacy.
{"title":"Fault-tolerant control for output synchronization of multi-agent systems: A data-driven approach","authors":"Wenli Chen , Xiaojian Li","doi":"10.1016/j.ins.2026.123127","DOIUrl":"10.1016/j.ins.2026.123127","url":null,"abstract":"<div><div>This study investigates the problem of data-driven fault-tolerant output synchronization control for heterogeneous multi-agent systems with unknown dynamics. Unlike the existing approaches that rely on prior knowledge of system matrices, this work proposes a novel method to design distributed data-driven fault-tolerant output synchronization controllers using data and output regulation theory. The output regulator equations are essential for fault-tolerant output synchronization control, whereas exact solutions from noise-corrupted data are challenging to obtain. To address this issue, a data-driven optimization problem is formulated to seek approximate solutions by minimizing the output regulation error matrices. Stability conditions are then derived in the form of data-dependent programs, whose solutions directly yield stabilizing feedback gains for agents. This approach ensures the achievement of output synchronization by utilizing data. Furthermore, a data-driven fault-tolerant controller is constructed by integrating adaptive control techniques with approximate solutions to the output regulator equations and stabilizing feedback gains learned from data, equipping agents with fault-tolerant capabilities. Theoretical analysis demonstrates that the proposed controller ensures the output synchronization errors are globally ultimately bounded (GUB). To validate the theoretical results, simulation examples are presented to demonstrate their efficacy.</div></div>","PeriodicalId":51063,"journal":{"name":"Information Sciences","volume":"738 ","pages":"Article 123127"},"PeriodicalIF":6.8,"publicationDate":"2026-01-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146081298","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}
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号