Neurocomputing最新文献

{"title":"Towards robust multimodal fault diagnosis of electromechanical systems with limited labeled data via cross-modal self-contrastive learning","authors":"Gaowei Xu ,&nbsp;Zian Lu ,&nbsp;Min Liu","doi":"10.1016/j.neucom.2026.133806","DOIUrl":"10.1016/j.neucom.2026.133806","url":null,"abstract":"<div><div>Multimodal signals with comprehensive and complementary information have been successfully applied in fault diagnosis of electromechanical systems. However, the scarcity of labeled multimodal signal data, coupled with inevitable distribution shifts, poses a significant challenge to the effective training of multimodal diagnostic models. Moreover, the availability of all modalities cannot always be guaranteed during extended inference periods, which can further induce significant performance degradation. Therefore, this paper proposes a robust multimodal fault diagnosis method with limited labeled data via a cross-modal self-contrastive learning (CMSCL) model. First, heterogeneous multimodal signals are collected and preprocessed to extract unified fault characteristics across different modalities. Then, the CMSCL model is initially pre-trained through modality-masking self-supervised learning on unlabeled signal data and subsequently fine-tuned with limited labeled data for fault diagnosis. Finally, a missing modalities completion module is designed and integrated into the CMSCL model to further address the missing modality issue. Extensive experimental results on two public experimental rig datasets and a real-world industrial dataset from different electromechanical systems demonstrate the superior accuracy and robustness of the proposed method compared with state-of-the-art approaches.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"691 ","pages":"Article 133806"},"PeriodicalIF":6.5,"publicationDate":"2026-05-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147827481","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":"HMP-Net: A hierarchical multi-prior network for brain tumor segmentation integrating physics, topology, and tumor dynamics","authors":"Yutong Wang ,&nbsp;Zhongfeng Kang ,&nbsp;Jiaxue Yang ,&nbsp;Shantian Yang ,&nbsp;Qinghua Zhao ,&nbsp;Zichen Song","doi":"10.1016/j.neucom.2026.133827","DOIUrl":"10.1016/j.neucom.2026.133827","url":null,"abstract":"<div><div>Precise brain tumor segmentation is essential for reliable diagnosis, treatment planning, and clinical follow-up. Despite recent progress in deep learning, most existing models remain predominantly data-driven and lack mechanisms to incorporate fundamental domain knowledge, including the physics of MRI acquisition, tumor morphology, and the biological dynamics of tumor progression. To bridge this gap, we propose <strong>HMP-Net</strong>, a theory-guided hierarchical multi-prior network that explicitly embeds these principles into the feature learning process. HMP-Net integrates three complementary levels of prior knowledge: (1) a <em>shallow physical signal encoder</em> that models inter-modal coupling in multimodal MRI data, (2) a <em>mid-level topological analyzer</em> that extracts Betti number–based structural priors through differentiable approximations, and (3) a <em>deep tumor dynamics modeler</em> that solves reaction–diffusion equations to capture biologically plausible tumor growth patterns. Extensive experiments on the BraTS 2021 and BraTS 2018 benchmarks demonstrate that HMP-Net surpasses state-of-the-art approaches, achieving average Dice scores of 91.55% and 86.58%, respectively. Ablation studies further validate the contribution of each hierarchical prior and show that the learned parameters maintain clear physical interpretability. These results demonstrate that embedding multi-scale, domain-specific priors into deep architectures substantially enhances generalization, interpretability, and clinical relevance, offering a new paradigm for knowledge-driven medical image analysis. The code will be available at <span><span>https://github.com/kanglzu/hmp_net</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"691 ","pages":"Article 133827"},"PeriodicalIF":6.5,"publicationDate":"2026-05-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147827383","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":"Robust jointly sparse 2-dimensional projection fuzzy clustering with local manifold structure preservation","authors":"Wu Chengmao ,&nbsp;Fengchao Gong","doi":"10.1016/j.neucom.2026.132970","DOIUrl":"10.1016/j.neucom.2026.132970","url":null,"abstract":"<div><div>Dimensionality reduction clustering methods combine feature reduction and clustering to analyze high-dimensional image data. However, 1D projection subspace clustering vectorizes 2D images into 1D vectors, disrupting spatial correlations and causing information loss. Two-stage models that separate reduction and clustering lack coordination, leading to suboptimal results. We propose a robust sparse two-dimensional projection fuzzy clustering method with local manifold constraints to improve image clustering. Each cluster is represented by a bilinear orthogonal subspace, and F1-norm reconstruction error updates sample memberships. A similarity matrix captures affinities, while a Laplacian matrix preserves manifold geometry during dimensionality reduction. Optimization uses block coordinate descent to alternately refine the projection matrix, cluster centroids, and membership matrix until convergence. This unified, unsupervised model avoids image vectorization, reducing computational complexity and preserving spatial relationships. Experiments on nine benchmark datasets show the RS2DPFC-LMS algorithm improves accuracy by 2.47 % and normalized mutual information by 2 %, demonstrating superior clustering performance, parameter stability, and noise robustness.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"676 ","pages":"Article 132970"},"PeriodicalIF":6.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146173297","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":"Entropy-aware event-triggered neural control for finite-time practical consensus of heterogeneous multi-agent systems under DoS attacks","authors":"Hongwei Ren ,&nbsp;Weiyi Li ,&nbsp;Zhiping Peng ,&nbsp;Feiqi Deng","doi":"10.1016/j.neucom.2026.133020","DOIUrl":"10.1016/j.neucom.2026.133020","url":null,"abstract":"<div><div>This paper investigates the finite-time practical consensus problem for heterogeneous second-order multi-agent systems subject to denial-of-service attacks. An entropy-aware event-triggered neural control framework is proposed that integrates multidimensional entropy-based attack detection across temporal, spatial, and frequency domains, entropy-guided adaptive event-triggering mechanisms, and finite-time control augmented by radial basis function neural network compensation for unknown heterogeneous dynamics. Rigorous Lyapunov-based theoretical analysis establishes finite-time practical consensus with explicit settling-time bounds dependent on initial conditions while excluding Zeno behavior. Simulation results demonstrate that, under diverse attack patterns, the proposed method achieves consensus in 10.04 s (4.0% faster than resilient event-triggered control) with only 4672 transmissions (approximately 80.5% reduction), validating superior attack resilience and communication efficiency.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"676 ","pages":"Article 133020"},"PeriodicalIF":6.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147386321","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":"EHC k-NN: Elliptic hypercomplex distance metrics for dimension-adaptive k-nearest neighbor","authors":"Kaan Arik ,&nbsp;Arzu Sürekçi ,&nbsp;Hidayet Hüda Kösal","doi":"10.1016/j.neucom.2026.133051","DOIUrl":"10.1016/j.neucom.2026.133051","url":null,"abstract":"&lt;div&gt;&lt;div&gt;This study introduces a dimension-adaptive k-Nearest Neighbor (k-NN) model that employs a family of elliptic hypercomplex distance metrics, addressing the limitations of Euclidean geometry in heterogeneous and correlated data with tabular and image datasets. The approach reshapes the feature space using a negative real parameter &lt;span&gt;&lt;math&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo&gt;&lt;&lt;/mo&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;/math&gt;&lt;/span&gt;, enabling curvature-controlled neighborhoods that better capture local structure. In the proposed method, each data instance is represented as an &lt;span&gt;&lt;math&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/math&gt;&lt;/span&gt;-dimensional elliptic hypercomplex number, and distances are computed through a norm that re-weights even- and odd- indexed components depending on &lt;span&gt;&lt;math&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;/math&gt;&lt;/span&gt;. The proposed method is dimension-adaptive in the sense that each real-valued feature vector of length &lt;span&gt;&lt;math&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/math&gt;&lt;/span&gt; is mapped to the smallest elliptic hypercomplex algebra of dimension &lt;span&gt;&lt;math&gt;&lt;msup&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mrow&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/math&gt;&lt;/span&gt; satisfying &lt;span&gt;&lt;math&gt;&lt;msup&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mrow&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;mo&gt;&lt;&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo&gt;≤&lt;/mo&gt;&lt;msup&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mrow&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/math&gt;&lt;/span&gt;. When &lt;span&gt;&lt;math&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo&gt;≠&lt;/mo&gt;&lt;msup&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mrow&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/math&gt;&lt;/span&gt;, the remaining components are zero-padded, so distance computations are carried out consistently in the corresponding &lt;span&gt;&lt;math&gt;&lt;msup&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mrow&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/math&gt;&lt;/span&gt;-dimensional elliptic hypercomplex space. Experiments were conducted on five tabular UCI + two image-derived benchmarks selected for their diversity in feature types and class structure. Performance was evaluated using classification performance evaluation metrics under identical &lt;span&gt;&lt;math&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/math&gt;&lt;/span&gt; settings. The proposed metric yields clear gains over Euclidean k-NN, particularly in &lt;em&gt;Wine&lt;/em&gt; (approximately &lt;span&gt;&lt;math&gt;&lt;mn&gt;2.0&lt;/mn&gt;&lt;/math&gt;&lt;/span&gt;-&lt;span&gt;&lt;math&gt;&lt;mn&gt;2.3&lt;/mn&gt;&lt;mi&gt;%&lt;/mi&gt;&lt;/math&gt;&lt;/span&gt;) and &lt;em&gt;Breast Cancer&lt;/em&gt; (approximately &lt;span&gt;&lt;math&gt;&lt;mn&gt;1.4&lt;/mn&gt;&lt;mi&gt;%&lt;/mi&gt;&lt;/math&gt;&lt;/span&gt;). Improvements are moderate in &lt;em&gt;Car Evaluation&lt;/em&gt;, while &lt;em&gt;Iris&lt;/em&gt; and &lt;em&gt;Banknote Authentication&lt;/em&gt; exhibit minimal change due to saturated separability and dominant attributes. On image-derived benchmarks (Seeds/Wheat and Image Segmentation), the proposed metric also delivers consistent improvements, typically around +2.0-2.7% in accuracy and +2.4-2.9% in F1-score compared with Euclidean k-NN. Further comparisons against metric-learning and manifold-inspired baselines (LMNN and geodesic distance) indicate that the proposed hypercomplex metric remains competitive and stable across neighborhood sizes, reinforcing its robustness beyond Euclidean geometry. Overall, the results indicate that the performance gains stem from the &lt;span&gt;&lt;math&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;/math&gt;&lt;/span&gt;-induced anisotropy of the elliptic hypercomplex norm, which reshapes neighborhood ","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"676 ","pages":"Article 133051"},"PeriodicalIF":6.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147386372","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":"Congestion-aware platoon re-sequencing optimization for electric vehicles using deep reinforcement learning","authors":"Chu Peng,&nbsp;Shaopan Guo,&nbsp;Miao Liu,&nbsp;Long Xiao","doi":"10.1016/j.neucom.2026.133040","DOIUrl":"10.1016/j.neucom.2026.133040","url":null,"abstract":"<div><div>With the development of Vehicle-to-Vehicle (V2V) communication, the non-fixed platoon method has become feasible, enabling vehicles to adjust positions dynamically, balance energy use, and improve efficiency. However, existing methods ignore the dynamic nature of traffic conditions. When road space is limited, platoon re-sequencing may become unsafe or even infeasible. To address these challenges, we propose a congestion-aware platoon re-sequencing optimization framework for electric vehicles (EVs) using deep reinforcement learning. The framework consists of two modules: a Traffic Congestion-Aware (TCA) module and a Deep Reinforcement Learning (DRL) module. Specifically, the TCA module predicts traffic congestion categories and incorporates them as constraints in the optimization process, overcoming the limitations of non-fixed platoon methods that neglect the safety and feasibility impacts of traffic congestion on re-sequencing. The DRL module, built on the Trust Region Policy Optimization (TRPO) algorithm, takes the EV State-of-Charge (SoC) and predicted traffic congestion categories as environmental observations. It restricts re-sequencing operations under congested conditions to prevent invalid actions and simultaneously manages the computational complexity that arises with increasing platoon size. Experimental results demonstrate that, compared to existing reinforcement learning methods without congestion constraints, our proposed framework reduces the frequency of platoon re-sequencing by 34.4%. Moreover, it achieves a 23.6% reduction in the final standard deviation of the SoC across all vehicles compared to existing re-sequencing algorithms, indicating that the unbalanced energy consumption of the vehicles has been reduced.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"676 ","pages":"Article 133040"},"PeriodicalIF":6.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147386361","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":"Automatic self-supervised learning for social recommendations","authors":"Xin He ,&nbsp;Wenqi Fan ,&nbsp;Ying Wang ,&nbsp;Mingchen Sun ,&nbsp;Xin Wang","doi":"10.1016/j.neucom.2026.133024","DOIUrl":"10.1016/j.neucom.2026.133024","url":null,"abstract":"<div><div>In recent years, researchers have leveraged social relations to enhance recommendation performance. However, most existing social recommendation methods require carefully designed auxiliary social tasks tailored to specific scenarios, which depend heavily on domain knowledge and expertise. To address this limitation, we propose Automatic Self-supervised Learning for Social Recommendations (AusRec), which integrates multiple self-supervised auxiliary tasks with an automatic weighting mechanism to adaptively balance their contributions through a meta-learning optimization framework. This design enables the model to automatically learn the optimal importance of each auxiliary task, thereby enhancing representation learning in social recommendations. Extensive experiments on several real-world datasets demonstrate that AusRec consistently outperforms state-of-the-art baselines by 3.3%–10.7% in Recall@10 and 1.4%–7.1% in NDCG@10, validating its effectiveness and robustness across different recommendation scenarios. The code is available at: <span><span>https://github.com/hexin5515/AusRec</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"676 ","pages":"Article 133024"},"PeriodicalIF":6.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147386376","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":"Adaptive wavelet decomposition and event-aware high-frequency modeling network for multivariate time series forecasting","authors":"Mengdi Gong,&nbsp;Chengci Wang,&nbsp;Jie Yu,&nbsp;Lingyu Xu","doi":"10.1016/j.neucom.2026.133089","DOIUrl":"10.1016/j.neucom.2026.133089","url":null,"abstract":"<div><div>Multivariate time series (MTS) forecasting has wide applications in real-world domains such as traffic, weather, and ocean monitoring. However, real-world MTS data often exhibit multiscale non-stationary patterns. These patterns arise from the dynamic coupling between long-term trends and sudden local events, making accurate forecasting highly challenging. Existing methods primarily rely on global modeling or fixed decomposition strategies. However, such approaches fail to adapt to varying spatiotemporal data and diverse task contexts. They cannot effectively disentangle trend and event sequences in accordance with the characteristics of time series data. Moreover, they cannot model unstable high-frequency events. To address these issues, we propose an Adaptive Wavelet Decomposition and Event-aware High-frequency Modeling Network (AweHF). The model employs an Adaptive Wavelet Decomposition module (AWD) to decouple the original sequence into low-frequency trends and high-frequency events in a data-driven way, avoiding the limitations of fixed wavelet bases. Subsequently, we apply a lightweight multilayer perceptron (MLP) to capture long-term dependencies in the trend component. In addition, we design a Time Aggregation Network (TAN) and Dual-Source Personalized Graph Convolution (DSPGC) to jointly model the volatility and instability of the event component. Finally, the bidirectional interaction fusion mechanism is used to integrate the trend and event components to fully exploit their complementary advantages. We conducted extensive experiments on six real-world datasets from multiple domains, and our results demonstrate that AweHF consistently outperforms all state-of-the-art baselines, achieving an average MAE reduction of more than 3.8% across the datasets. Code is available at this repository: <span><span>https://github.com/WangChengci/AweHF</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"676 ","pages":"Article 133089"},"PeriodicalIF":6.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147386319","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":"Time-frequency-based pyramid channel network for long-term time series forecasting","authors":"Zhiqiang Jiang ,&nbsp;Yongsheng Dong ,&nbsp;Min Han ,&nbsp;Haotian Yang ,&nbsp;Xiaotong Chen","doi":"10.1016/j.neucom.2026.133022","DOIUrl":"10.1016/j.neucom.2026.133022","url":null,"abstract":"<div><div>Many time-domain and frequency-domain based methods have been proposed for long-term time series forecasting. In order to obtain the seasonal correlation of different channels and time series features at different time scales, we propose a brand-new time-frequency-based pyramid channel network (TPCNet) for long-term time series forecasting. Particularly, we first build a multi-channel seasonal feature attention residual fusion structure to obtain seasonal correlations between different channels by using the short-time Fourier transform, residual ideas, and fusion operations of multiple kernels’ different channels. We then propose a dual-dimensional attention residual pyramid structure to obtain time series features at different time scales by using tensor summation operations, residual ideas, and attention mechanisms. Finally, we obtain time-series prediction results through fully connected operations. Our proposed TPCNet shows competitive prediction performance when compared with many sample classical methods on GeForce RTX 4060Ti, according to the results of experiments on six commonly used time series datasets.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"676 ","pages":"Article 133022"},"PeriodicalIF":6.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147386370","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":"Fight light with light: A review of physical adversarial attack within light transmission pipeline","authors":"Guojia Li ,&nbsp;Simin Xu ,&nbsp;Yan Cao ,&nbsp;Mingyue Cao ,&nbsp;Yihong Zhang","doi":"10.1016/j.neucom.2026.133034","DOIUrl":"10.1016/j.neucom.2026.133034","url":null,"abstract":"<div><div>Deep Neural Networks (DNNs) remain vulnerable to physical adversarial attacks. Attacks that target the light transmission pipeline exhibit heightened stealthiness while posing severe real-world threats due to their flexible and deployable nature. To advance the understanding of this emerging threat, we establish a unified framework that systematically analyzes the entire light transmission pipeline as a contiguous attack surface. Within this framework, we identify two primary attack vectors, manipulating light transmission channel and attacking image perception device, and systematically characterize their methodologies across nine key attributes. We further formalize the optimization process for generating adversarial light patterns and assess the physical deployment methods of such attacks. Furthermore, we propose a graded framework for evaluating the transferability and demonstrate that while physical adversarial examples in this domain exhibit high stealthiness, their transferability across different model architectures remains limited. Finally, we outline current challenges and discuss future research directions.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"676 ","pages":"Article 133034"},"PeriodicalIF":6.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147386328","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}