Neurocomputing最新文献

{"title":"AttackTracer: Semantic-level adversarial attack location traceability via evidential diffusion model","authors":"Zhentong Zhang ,&nbsp;Xinde Li ,&nbsp;Pengfei Zhang ,&nbsp;Kui Wang ,&nbsp;Tianrong Gao ,&nbsp;Tao Shen","doi":"10.1016/j.neucom.2025.131535","DOIUrl":"10.1016/j.neucom.2025.131535","url":null,"abstract":"<div><div>Adversarial attacks pose a significant threat to AI systems, yet existing detection methods mainly focus on image-level threats, limiting fine-grained localization of perturbations. To address this challenge, we propose AttackTracer, the first semantic-level localization framework specifically designed for instance-level adversarial attacks. Instance-level adversarial perturbations are typically sparse and localized, which aligns naturally with the capabilities of diffusion models to progressively reconstruct sparse structures from stochastic noise. Building on this property, AttackTracer models the adversarial mask as a conditional distribution given the adversarial image, allowing iterative refinement and effective recovery of attack regions. To address the inherent instability of diffusion sampling, we introduce the Temporal Evidence Fusion Strategy (TEFS). TEFS integrates Dempster–Shafer theory with a signal-to-noise-ratio (SNR)-guided temporal ensemble, aggregating multi-step predictions to mitigate conflicts and uncertainty, thus achieving robust inference. Furthermore, adversarial perturbations often manifest as subtle high-frequency and edge distortions. To capture these, AttackTracer employs two complementary modules: the Wavelet Frequency Fusion Block (WFFB), which extracts multi-scale frequency features via Discrete Wavelet Transform to enhance sensitivity to sparse perturbations, and the Edge Feature Enhancement Module (EFEM), which models multi-granularity edge structures using parallel branches and FFT to detect boundary distortions. Together, WFFB and EFEM provide complementary views of perturbation patterns. Extensive experiments demonstrate that AttackTracer achieves superior traceability of adversarial regions while maintaining robustness across stochastic sampling and varying scales, highlighting its effectiveness for instance-level attack localization.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"656 ","pages":"Article 131535"},"PeriodicalIF":6.5,"publicationDate":"2025-09-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145119823","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Feature selection via risk-bound utility maximization","authors":"Chunxu Cao ,&nbsp;Qiang Zhang","doi":"10.1016/j.neucom.2025.131572","DOIUrl":"10.1016/j.neucom.2025.131572","url":null,"abstract":"<div><div>The ultimate goal of supervised feature selection is to identify a feature subset that minimizes classification risk. Contemporary methods, however, often rely on heuristic or model-dependent proxy criteria that lack a direct theoretical connection to this fundamental objective. To bridge this gap, we introduce a new feature selection framework that directly optimizes a model-agnostic utility function grounded in statistical learning theory. Our approach defines the utility of a feature subset based on the 1-Wasserstein distance between class-conditional distributions. This metric is theoretically powerful as it can be used to construct an upper bound on the Bayes classification error, allowing us to construct a utility function that is a direct surrogate for this risk bound. We instantiate this framework with a subset search strategy that effectively captures feature interactions by maximizing this risk-bound utility. Extensive experiments on real-world datasets demonstrate that our method not only achieves state-of-the-art classification performance but also demonstrates superior robustness and interpretability, providing a principled and powerful alternative to traditional feature selection methods, confirming our framework’s theoretical soundness.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"657 ","pages":"Article 131572"},"PeriodicalIF":6.5,"publicationDate":"2025-09-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145157037","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A zero-shot high-performance fire detection framework based on large language models","authors":"Hongyang Zhao ,&nbsp;Yi Liu ,&nbsp;Yuhang Han ,&nbsp;Xingdong Li ,&nbsp;Yanan Guo ,&nbsp;Jing Jin","doi":"10.1016/j.neucom.2025.131403","DOIUrl":"10.1016/j.neucom.2025.131403","url":null,"abstract":"<div><div>Fire detection is crucial for minimizing economic damage and safeguarding human lives. Existing methods, including advanced AI and ML techniques, face challenges such as detecting small fires in complex environments and relying on extensive labeled data for training. This paper proposes a novel zero-shot fire detection framework leveraging large language models (LLMs) and contrastive learning-based image–text pre-training models. The framework introduces an enhanced self-attention mechanism for optimizing image embeddings, diverse prompt generation using GPT-3.5 for improved generalization, and a dynamic threshold calculation method based on statistical analysis to enhance detection accuracy and reliability. The proposed method is tested on the public FLAME dataset and a self-collected dataset. Experimental results demonstrate that the proposed method outperforms state-of-the-art models in detecting small fires within complex backgrounds, achieving better detection performance without the need for any training data. This study highlights the potential of zero-shot learning in fire detection and provides a promising solution for real-world fire detection applications.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"657 ","pages":"Article 131403"},"PeriodicalIF":6.5,"publicationDate":"2025-09-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145156650","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Double missing multi-view multi-label classification via an attention-guided multi-space consistency alignment framework","authors":"Bingyan Nie ,&nbsp;Wulin Xie,&nbsp;Lian Zhao,&nbsp;Jiang Long,&nbsp;Xiaohuan Lu,&nbsp;Yinghao Ye","doi":"10.1016/j.neucom.2025.131558","DOIUrl":"10.1016/j.neucom.2025.131558","url":null,"abstract":"<div><div>Multi-view multi-label classification (MVMLC) seeks to enhance classification by integrating diverse data views, but its practical use is hindered by missing views and labels, posing the significant challenge of incomplete MVMLC(IMVMLC). Although various IMVMLC approaches have been proposed, most of them handle multiple objectives in a single feature space and thus overlook the conflict between learning consistent common semantics and reconstructing view-specific information. In addition, existing multi-view classification methods mainly consider utilizing the features of each view, while ignoring the inconsistent contributions of each view and usually relying on static average weighting strategies. To this end, we propose our Attention-Guided MultiSpace Consistency Alignment Framework (AMCA). In Stage 1, AMCA introduces multi-space representation learning with dual-level contrastive objectives, explicitly disentangling shared and view-specific semantics to resolve the objective conflict and yield more informative embeddings. In Stage 2, AMCA employs an attention-guided fusion module that dynamically evaluates and integrates multi-view features based on their relevance to the classification task, enabling robust decision-making even with missing data. Extensive experiments validate the effectiveness and superiority of our proposal.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"656 ","pages":"Article 131558"},"PeriodicalIF":6.5,"publicationDate":"2025-09-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145107486","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Constrained multi-agent evasion using deep reinforcement learning","authors":"Bowei Yan ,&nbsp;Runle Du ,&nbsp;Xiaojun Ban ,&nbsp;Di Zhou","doi":"10.1016/j.neucom.2025.131550","DOIUrl":"10.1016/j.neucom.2025.131550","url":null,"abstract":"<div><div>Designing effective evasion strategies in pursuit–evasion scenarios is challenging, particularly when the pursuer’s model is unknown and inaccessible. This limitation hinders the application of conventional evasion policy design methods. To overcome this challenge, especially when evaders have constrained maneuverability against unrestricted pursuers, we propose a novel multi-agent evasion algorithm based on deep reinforcement learning. Our approach employs a staged learning framework, progressively guiding evaders from simpler to more complex tasks to refine their evasion strategies. Crucially, our algorithm enables evaders to infer pursuers’ intentions even without prior knowledge of pursuers’ objectives, allowing for optimal decision-making despite mobility constraints. Simulation results demonstrate that our method significantly enhances evasion success, validating the effectiveness of learning-based strategies. Additionally, the algorithm exhibits strong adaptability to environmental changes, ensuring reliable performance across diverse pursuit–evasion scenarios.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"657 ","pages":"Article 131550"},"PeriodicalIF":6.5,"publicationDate":"2025-09-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145120147","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Frequency-aware fusion for improved video object segmentation","authors":"Zhiqiang Hou ,&nbsp;Hao Cui ,&nbsp;Chenxu Wang ,&nbsp;Sugang Ma ,&nbsp;Xiaobao Yang ,&nbsp;Lei Pu","doi":"10.1016/j.neucom.2025.131585","DOIUrl":"10.1016/j.neucom.2025.131585","url":null,"abstract":"<div><div>Currently, most mainstream memory-based semi-supervised video object segmentation (VOS) methods rely on pixel-level matching to identify target objects. However, the majority of these approaches depend solely on spatial-domain features for representation, which limits their ability to preserve fine-grained details. In addition, they typically adopt a single bottom-up matching strategy, which lacks sufficient global semantic guidance, ultimately leading to suboptimal segmentation performance. To address these issues, we propose a Frequency-Aware Fusion for Improved Video Object Segmentation algorithm (FAFVOS), which incorporates frequency-domain information enhancement and a bidirectional matching mechanism to improve segmentation accuracy. First, we design a Hierarchical Frequency-Aware Encoder (HFAE), which enhances shallow features by leveraging high-frequency components to preserve edge and texture details, and strengthens deep features via low-frequency components to maintain global structural consistency, thereby achieving multi-scale frequency–spatial feature fusion. Second, a frequency-guided bidirectional matching Transformer module is proposed to establish pixel-level and object-level dual-path interactions. By incorporating a cross-attention mechanism, the model effectively facilitates joint reasoning between local pixel-wise details and global object-level semantics. Finally, a high-order moment refinement module is introduced to integrate high-order statistical features, enhancing the model’s ability to capture object deformation and leading to high-quality segmentation results. The proposed method is evaluated on the DAVIS, YouTube-VOS, and MOSE datasets. Experimental results demonstrate that, without relying on complex pretraining strategies or additional datasets, our approach achieves a real-time inference speed of 56 FPS with a <span><math><mrow><mi>J</mi></mrow><mi>&amp;</mi><mrow><mi>F</mi></mrow></math></span> score of 88.5 % on the DAVIS 2017 benchmark, surpassing existing representative methods. Moreover, it also achieves consistently superior performance on the more challenging YouTube-VOS and MOSE datasets, further validating the generalization ability and robustness of the proposed approach.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"656 ","pages":"Article 131585"},"PeriodicalIF":6.5,"publicationDate":"2025-09-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145119991","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Single-client GAN-based backdoor attacks for Asynchronous Federated Learning","authors":"Siyu Guan,&nbsp;Chunguang Huang,&nbsp;Hai Cheng","doi":"10.1016/j.neucom.2025.131580","DOIUrl":"10.1016/j.neucom.2025.131580","url":null,"abstract":"<div><div>Federated Learning (FL) enables distributed collaborative training while preserving data privacy; however, it demonstrates significant vulnerability to backdoor attacks. Existing attack methodologies predominantly require control of numerous malicious clients to achieve efficacy and largely neglect asynchronous FL scenarios. In response to these limitations, we propose a novel GAN-based backdoor attack framework capable of injecting effective and covert backdoors with minimal malicious client participation, functioning efficiently across both synchronous and asynchronous environments. Our framework operates effectively with a single malicious client, eliminating the need for coordination among multiple adversarial participants or prior knowledge of benign client data distributions. This reduction in resource requirements enhances the framework's practicality in real-world FL implementations. The malicious client employs a Generative Adversarial Network to synthesize adversarial samples containing predefined triggers, which are subsequently incorporated into local training datasets. The concurrent training on legitimate and triggered data enhances attack effectiveness, while gradient injection—manipulating differences between local and global gradients to introduce strategic noise—facilitates backdoor embedding with improved stealth characteristics. Empirical evaluations demonstrate that in a configuration of 200 clients with a single attacker, our framework achieves attack success rates of 98.66 % on MNIST and 86.29 % on CIFAR-10 datasets. Comprehensive experimentation across both datasets substantiates the framework's effectiveness, imperceptibility, and resilience in synchronous and asynchronous FL environments. This research contributes significant insights into backdoor attack strategies in FL, particularly within asynchronous contexts, and underscores the imperative for developing robust defensive countermeasures.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"657 ","pages":"Article 131580"},"PeriodicalIF":6.5,"publicationDate":"2025-09-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145120138","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"WaterBox: Weakly supervised underwater instance segmentation and a new benchmark","authors":"Meng Wu ,&nbsp;Yifeng Cui ,&nbsp;Rong Min ,&nbsp;Shanghang Jiang ,&nbsp;Lei Zhang ,&nbsp;Jing Yu","doi":"10.1016/j.neucom.2025.131582","DOIUrl":"10.1016/j.neucom.2025.131582","url":null,"abstract":"<div><div>Box-supervised instance segmentation has gained increasing attention due to its reliance on weak box annotations, which are considerably less expensive than pixel-wise mask annotations. Despite the advantage, existing methods in this category often struggle in complex underwater scenes, where degraded image quality causes foreground objects to become heavily entangled with the background. To address this issue, we propose WaterBox, a cost-effective box-supervised underwater instance segmentation method. Considering the intrinsic characteristics of underwater imaging, we introduce a novel pairwise loss function that leverages a mixed color affinity map with a dynamic threshold to effectively disambiguate foreground and background boundaries. Additionally, we devise a bounding box refinement strategy that generates tight and accurate bounding boxes for each instance, alleviating the negative impact of imprecise box annotations on segmentation performance. Furthermore, to fill in the gaps caused by data scarcity, we construct the first diver instance segmentation dataset, DSeg, which consists of 2000 underwater images with high-quality instance masks. Extensive experiments on two underwater datasets demonstrate the superiority of our approach over the state-of-the-art (SOTA) weakly supervised methods. The code and dataset will be made publicly available.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"657 ","pages":"Article 131582"},"PeriodicalIF":6.5,"publicationDate":"2025-09-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145223329","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Scattered data augmentation for generalization in visual reinforcement learning","authors":"Hao Lei,&nbsp;Yu Zhao,&nbsp;Yi Xin,&nbsp;Zhang Shaonan,&nbsp;Ke Liangjun","doi":"10.1016/j.neucom.2025.131492","DOIUrl":"10.1016/j.neucom.2025.131492","url":null,"abstract":"<div><div>Data augmentation (DA) has shown a significant potential to enhance generalization performance in visual reinforcement learning (VRL). However, existing research on DA-based methods is predominantly empirical, and the mechanism for why DA enhances generalization remains theoretically under-explored. To bridge this gap, we derive a generalization error upper bound for VRL from the perspective of data distribution distance. Based on this bound, we provide a theoretical explanation of the mechanism by which DA improves generalization: we find that DA that satisfies certain conditions can reduce the distance between the training and test distributions, thus making the training and test samples closer. In addition, we conditionally prove that training data with higher variance can provide a higher generalization performance. Motivated by our analysis, we propose Scattered Data Augmentation (ScDA) framework. ScDA constructs a data transformation system with the agent serving as the discriminator, aiming to provide more diverse training data for agent training. Experiments are conducted across various tasks and numerous test modes in DeepMind Control Generalization Benchmark2 (DMC-GB2) and robotic tasks. Results demonstrate that our ScDA framework can be integrated with different baseline algorithms and significantly enhance policy generalization, outperforming the current state-of-the-art methods in the DMC-GB2 tests, confirming the effectiveness of the theoretical analysis in this work. The code for this work can be found at: <span><span>https://github.com/scdadev/scdadev</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"656 ","pages":"Article 131492"},"PeriodicalIF":6.5,"publicationDate":"2025-09-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145108160","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Urban multi-domain mixing (UMDMix) based unsupervised domain adaptation for LiDAR semantic segmentation","authors":"Anurag Nihal ,&nbsp;Pyare Lal ,&nbsp;Vaibhav Kumar","doi":"10.1016/j.neucom.2025.131526","DOIUrl":"10.1016/j.neucom.2025.131526","url":null,"abstract":"<div><div>3D semantic maps generated from Light Detection and Ranging (LiDAR) point clouds enable scene understanding in diverse applications such as autonomous driving and urban planning. However, existing deep learning models struggle when tested on different domains, worsened by limited labeled data. Unsupervised Domain Adaptation (UDA) can bridge this gap, but existing UDA methods often face adaptation challenges due to domain shifts arising from variations in the physical environment, data sparsity, and sensor differences. To address these limitations, we propose <em>UMDMix</em>, a novel UDA architecture that operates on the mixing of multiple labeled source domains with unlabeled target domains to make the predictive model robust to cross-domain variations. <em>UMDMix</em> integrates a teacher–student learning scheme to produce a robust teacher model and an adaptable student model. The performance of the teacher model in the source domain is further strengthened by a position-aware loss that assigns greater significance to semantically rich neighborhoods. A combination of entropy regularization and KL-divergence loss in the target domain updates the knowledge of the teacher model to the student model during adaptation. Our extensive experiments across diverse environments show that <em>UMDMix</em> achieves an average improvement of 13 % on minor classes such as bicycle, traffic sign, and person in target domain datasets, outperforming previous State-Of-The-Art (SOTA) UDA methods.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"657 ","pages":"Article 131526"},"PeriodicalIF":6.5,"publicationDate":"2025-09-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145157040","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}