Information Fusion最新文献

{"title":"An Audio-Visual Fusion Emotion Generation Model Based on Neuroanatomical Alignment","authors":"Sir Haidong Wang, Qia Shan, JianHua Zhang, PengFei Xiao, Ao Liu","doi":"10.1016/j.inffus.2026.104739","DOIUrl":"https://doi.org/10.1016/j.inffus.2026.104739","url":null,"abstract":"In the field of affective computing, traditional methods for generating emotions predominantly rely on deep learning techniques and large-scale emotion datasets. However, deep learning techniques are often complex and difficult to interpret, and standardizing large-scale emotional datasets are difficult and costly to establish. To tackle these challenges, we introduce a novel framework named Audio-Visual Fusion for Brain-like Emotion Learning(AVF-BEL). In contrast to conventional brain-inspired emotion learning methods, this approach improves the audio-visual emotion fusion and generation model through the integration of modular components, thereby enabling more lightweight and interpretable emotion learning and generation processes. The framework simulates the integration of the visual, auditory, and emotional pathways of the brain, optimizes the fusion of emotional features across visual and auditory modalities, and improves upon the traditional Brain Emotional Learning (BEL) model. The experimental results indicate a significant improvement in the similarity of the audio-visual fusion emotion learning generation model compared to single-modality visual and auditory emotion learning and generation model. Ultimately, this aligns with the fundamental phenomenon of heightened emotion generation facilitated by the integrated impact of visual and auditory stimuli. This contribution not only enhances the interpretability and efficiency of affective intelligence but also provides new insights and pathways for advancing affective computing technology. Our source code can be accessed here: <ce:inter-ref xlink:href=\"https://github.com/OpenHUTB/emotion\" xlink:type=\"simple\">https://github.com/OpenHUTB/emotion.</ce:inter-ref>","PeriodicalId":50367,"journal":{"name":"Information Fusion","volume":"69 1","pages":""},"PeriodicalIF":18.6,"publicationDate":"2026-08-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148884305","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":"ESV-FNet: Multi-Neural-View and Multimodal Evidence Fusion for Occupant-Experienced Road-Rage-Related Negative Affect Recognition in Real-Road Traffic Conflicts","authors":"Yuchen Guo, Tianyao Zhang, Liyun Deng, Jiaxue Cai, Zhenhai Gao","doi":"10.1016/j.inffus.2026.104740","DOIUrl":"https://doi.org/10.1016/j.inffus.2026.104740","url":null,"abstract":"Negative affect experienced by vehicle occupants during real-road traffic conflicts is transient, subjective, and jointly shaped by external vehicle-motion stimuli and internal neural responses. Reliance on overt behavior or scalp electroencephalography (EEG) alone may therefore limit both recognition robustness and physiological interpretability. This study proposes an EEG-Source-Vehicle Fusion Network (ESV-FNet) for four-level recognition of traffic-conflict-induced road-rage-related negative affect experienced by front-seat occupants. The framework treats the template-based source estimates reconstructed from scalp EEG not as an independent sensor modality but as an approximate source-space representation of the same recorded neural signals. It integrates raw and source-space neural representations through shared-private disentanglement, bidirectional cross-attention, and source-contribution modulation; encodes vehicle speed, longitudinal acceleration, lateral acceleration, and yaw rate through a dual-branch vehicle-dynamics encoder; and fuses neural and vehicle evidence by modeling both cross-source discrepancy and agreement. A real-road experiment involving 22 participants seated in the front passenger seat was conducted under longitudinal threat and lateral intrusion scenarios controlled by a professional safety driver, yielding synchronized EEG, vehicle-dynamics, and subjective rating data. The operational labels represent self-reported occupant-experienced negative affect rather than active-driver road rage or aggressive driving behavior and may contain overlapping components of irritation, threat perception, fear, discomfort, stress, and general negative arousal. Under leave-one-subject-out validation, ESV-FNet achieved balanced accuracies of 82.70% and 84.71% in the two scenarios. Under a fixed train/validation/test split, it achieved overall accuracies of 95.28% and 97.25%. Ablation analyses further supported the modeling utility of including the source-space-derived representation and vehicle-dynamics context. However, because the source-space representation is deterministically derived from scalp EEG and is processed through an additional model branch, the observed improvement cannot be attributed to spatial reparameterization alone. These results suggest that structured fusion of homologous neural views and vehicle-motion evidence provides an effective framework for recognizing occupant-experienced road-rage-related negative affect within the present traffic-conflict paradigm.","PeriodicalId":50367,"journal":{"name":"Information Fusion","volume":"6 1","pages":""},"PeriodicalIF":18.6,"publicationDate":"2026-08-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148884122","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":"MCEF: Multiscale contrastive enhanced fusion network for hyperspectral and LiDAR data classification","authors":"Ziqi Li, Yu Yan, Yonghong Zhang, Tianyang Xu","doi":"10.1016/j.inffus.2026.104736","DOIUrl":"https://doi.org/10.1016/j.inffus.2026.104736","url":null,"abstract":"Multisensor data fusion enhances classification accuracy by integrating complementary information. A typical example is the fusion of hyperspectral imagery (HSI) and light detection and ranging (LiDAR) data, which has attracted considerable research interest due to their complementary spectral, spatial, and 3D structural attributes. However, prevailing methods typically rely on shallow fusion strategies, which fail to achieve deep cross modal interaction and struggle to capture complex semantic discrepancies and multiscale contextual dependencies. A multiscale contrastive enhanced fusion (MCEF) network is introduced herein to address these shortcomings, specifically designed for joint HSI-LiDAR classification. The MCEF architecture is centered on three key components: a cross modal difference enhancement (CMDE) module that amplifies cross modal discrepancies to highlight complementary information; a multiscale transformer fusion (MTF) module that aggregates local fine-grained details with global contextual cues; and a cross modal contrastive learning (CMCL) module that improves feature discriminability and cross modal consistency through semantic-aware constraints. Experimental results on three public benchmarks demonstrate that the proposed MCEF network achieves state-of-the-art performance, particularly in scenes with high structural and textural complexity. Additional ablation studies corroborate the contribution of each constituent module. The implementation code for the MCEF model is accessible at: <ce:inter-ref xlink:href=\"https://github.com/li-zi-qi/MCEF\" xlink:type=\"simple\">https://github.com/li-zi-qi/MCEF</ce:inter-ref>.","PeriodicalId":50367,"journal":{"name":"Information Fusion","volume":"15 1","pages":""},"PeriodicalIF":18.6,"publicationDate":"2026-08-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148884142","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":"NSAP: Neuro-symbolic adaptive propagation for knowledge graph completion","authors":"Xinde Yu, Hui Zhang, Bo Zhang, Kun Gao, Zhao Jin, Mingliang Xu","doi":"10.1016/j.inffus.2026.104737","DOIUrl":"https://doi.org/10.1016/j.inffus.2026.104737","url":null,"abstract":"Progressive propagation graph neural networks have emerged as a powerful paradigm for knowledge graph completion because of their ability to perform multi-hop reasoning. These methods typically conduct query-specific, layer-wise message propagation over dynamically constructed subgraphs, enabling representations to adapt to the target query. However, they still suffer from two major limitations. First, different orderings of the same set of relations may represent distinct semantic compositions, whereas existing propagation mechanisms often fail to preserve such ordering information, potentially conflating semantically different reasoning paths. Second, deeper propagation rapidly expands the set of involved entities, which dilutes informative signals and incurs substantial computational overhead. To address these limitations, we propose Neuro-Symbolic Adaptive Propagation (NSAP), a framework that incorporates Horn-rule priors into progressive message passing to guide reasoning along semantically consistent multi-hop trajectories. Specifically, NSAP maintains the matching progress of activated rules over the query-dependent subgraph and assigns each candidate edge a rule-consistency score according to its compatibility with the next expected relation in these rules. The resulting scores are incorporated into the attention mechanism to promote order-sensitive propagation and mitigate semantic drift. In addition, NSAP employs rule-aware adaptive node sampling, which prioritizes entities that are more likely to advance relevant rules while filtering out distracting entities, thereby limiting unnecessary subgraph expansion. Experiments on multiple benchmark knowledge graphs demonstrate that NSAP delivers consistent overall improvements over strong baselines in both transductive and inductive knowledge graph completion settings.","PeriodicalId":50367,"journal":{"name":"Information Fusion","volume":"76 1","pages":""},"PeriodicalIF":18.6,"publicationDate":"2026-08-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148884145","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":"MIGO: A mutual-information fusion framework for heterogeneous cross-modality generation in single-cell multi-omics","authors":"Xuanwei Lin, Yu Sun, Xiaotang Wang, Peng Zhang, Yaoyao Zhang, XueMeng Li, Yuyang Wang, Pengzhen Hu, Yongqi Zhang, Hebing Chen, Xiaochen Bo, Hao Li","doi":"10.1016/j.inffus.2026.104738","DOIUrl":"https://doi.org/10.1016/j.inffus.2026.104738","url":null,"abstract":"Mapping information flow along the central dogma across genome organization, chromatin accessibility, transcription, translation, and protein abundance is fundamental to understanding cellular heterogeneity and function. However, single-cell multi-omics datasets remain fragmented across assays, and most existing cross-modality methods focus on a limited range of vector-valued molecular readouts, making it difficult to connect heterogeneous modalities with distinct data geometries. Here, we present MIGO, a mutual-information-based fusion framework applied across modality-paired datasets spanning scRNA-seq, scATAC-seq, scHi-C, ribosome profiling, and protein abundance. MIGO separates biological-state representations from technical variation, aligns paired biological states, and organizes the aligned representation through vector quantization, while modality-specific decoders support both vector-valued and structured outputs. Across 20 dataset-level benchmarks, MIGO demonstrates strong cross-modal performance and batch-robust representation learning while preserving modality-specific biological structure. Notably, MIGO generates scHi-C contact maps from scRNA-seq and recovers chromatin contact-decay patterns and higher-order structural features. It also recovers cell-type-specific translational-efficiency programs and protein-abundance profiles that preserve marker identity and cellular organization. Together, MIGO provides a common information-fusion framework for transcriptome-centered generation across heterogeneous molecular readouts.","PeriodicalId":50367,"journal":{"name":"Information Fusion","volume":"24 1","pages":""},"PeriodicalIF":18.6,"publicationDate":"2026-08-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148884143","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":"Graph-based Representation for Specificity and Heterogeneity in Biomedical Multi-view Learning","authors":"Shuang Feng, Xiaohua Wan, Chang Li, Simeng Liu, Shuxian Liu, Leting Wang, Xiahai Zhuang, Fa Zhang, Bin Hu","doi":"10.1016/j.inffus.2026.104734","DOIUrl":"https://doi.org/10.1016/j.inffus.2026.104734","url":null,"abstract":"Biomedical multi-view learning (BioMVL) has shown promise for disease classification by integrating complementary information across views. However, existing methods often overlook the specificity of each view and the semantic heterogeneity across multi-view data, limiting representation quality and interpretability. To address these problems, this paper presents a novel specificity–heterogeneity BioMVL framework (SH-BioMVL). The framework employs a global–local graph neural network (GLGNN) to disentangle the interactivity and specificity information in multi-view data. Additionally, the LD-based individual genetic graph construction method (LDIG) transforms high-dimensional genetic sequences into biologically informed representations to facilitate heterogeneous fusion. Experimental validation across different cohorts, modalities, and diseases demonstrates that SH-BioMVL achieves state-of-the-art performance and maintains robust generalization in biomedical multi-view disease classification tasks. Analysis of cross-view interactivity and specificity features further confirms the biological validity and reliability. Overall, SH-BioMVL provides an effective framework for addressing specificity and heterogeneity in BioMVL, offering strong support for solving biomedical problems.","PeriodicalId":50367,"journal":{"name":"Information Fusion","volume":"43 1","pages":""},"PeriodicalIF":18.6,"publicationDate":"2026-08-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148884144","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":"Efficient federated learning framework for IoT intrusion detection: Two-stage hierarchical traffic classification method","authors":"Jingnan Dong, Haolei Chen, Shigen Shen, Guangxia Xu, Jun Liu, Zhiquan Liu","doi":"10.1016/j.inffus.2026.104730","DOIUrl":"https://doi.org/10.1016/j.inffus.2026.104730","url":null,"abstract":"The widespread adoption of Internet of Things technology has significantly increased its vulnerability to cyberattacks. Traditional intrusion detection systems rely on centralized data analysis, which is susceptible to privacy leakage and incurs high computational costs. As a decentralized machine learning paradigm, federated learning provides an effective means of maintaining data privacy while enabling collaborative training of robust models for network intrusion detection. However, existing federated learning-based methods still face challenges related to resource efficiency. For example, many existing single-stage federated multi-class intrusion detection methods treat benign traffic as a separate class. In federated IoT environments, where benign traffic overwhelmingly dominates local client datasets, this design introduces redundant computation and amplifies benign-biased gradients during aggregation, resulting in both efficiency degradation and weakened attack discrimination capability. To address these challenges, we propose a two-stage federated learning-based intrusion detection method, the Two-Stage Classifier for Federated Learning (TSC-FL). In the first stage, a ResNet-based binary traffic detector (ResBTD) performs binary classification to identify and isolate malicious traffic, thereby reducing computational overhead. In the second stage, a Mamba-based multi-class traffic detector (MamMTD) conducts fine-grained classification of the malicious traffic. The Mamba-based classifier is adopted to efficiently model long-range dependencies in traffic sequences with linear-time complexity. This strategy effectively alleviates the imbalance between benign and malicious categories, while simultaneously improving detection efficiency and model performance. Experiments on the UNSW-NB15, CICIDS2017, and N-BaIoT datasets demonstrate the effectiveness and efficiency of our methodology.","PeriodicalId":50367,"journal":{"name":"Information Fusion","volume":"135 1","pages":""},"PeriodicalIF":18.6,"publicationDate":"2026-08-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148884148","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":"Learning from incomplete signals: Multi-view knowledge distillation for robust harmful micro-Video detection","authors":"Zhi Zeng, Jiaying Wu, Xulang Zhang, Xiangzheng Kong, Herun Wan, Guodong Li, Qika Lin, Minnan Luo, Erik Cambria","doi":"10.1016/j.inffus.2026.104735","DOIUrl":"https://doi.org/10.1016/j.inffus.2026.104735","url":null,"abstract":"As hateful and misleading content continues to proliferate on social media platforms, detecting multimodal harmful content in micro-videos has become increasingly important. While recent advances in this area have shown promise, existing methods often assume complete access to all modalities, including video, audio, and textual descriptions. This assumption fundamentally limits their robustness in real-world settings, where micro-videos are frequently <ce:italic>modality- and segment-incomplete</ce:italic> due to upload artifacts, platform-level compression, or missing metadata. To bridge this gap, we introduce three real-world benchmarks that systematically capture varying degrees of incompleteness at both the modality and segment levels. We further propose <ce:bold>MVKD</ce:bold>, a multi-view knowledge distillation framework tailored for robust harmful micro-video detection under incomplete and uncertain conditions. MVKD (1) adaptively distills <ce:italic>modality-specific</ce:italic> knowledge through confidence-aware guidance, (2) enhances <ce:italic>sample- and domain-level</ce:italic> representations via contrastive alignment, and (3) integrates complementary cues using a <ce:italic>multi-view mixture-of-experts</ce:italic> strategy. Extensive experiments show that MVKD consistently outperforms strong baselines in both accuracy and robustness across diverse missingness patterns, and generalizes effectively to previously unseen domains.","PeriodicalId":50367,"journal":{"name":"Information Fusion","volume":"3 1","pages":""},"PeriodicalIF":18.6,"publicationDate":"2026-08-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148884146","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":"SecHFed: Achieving secure hierarchical federated learning against poisoning attacks without data assumptions","authors":"Chen Gu, Kun Zhu, Donghui Hu, Zhiquan Liu","doi":"10.1016/j.inffus.2026.104731","DOIUrl":"https://doi.org/10.1016/j.inffus.2026.104731","url":null,"abstract":"Federated Learning (FL) is a promising paradigm for distributed model training where a central server produces a global model by aggregating gradients from distributed clients while retaining data locally. Due to multi-tier structural requirements, Hierarchical Federated Learning (HFL) is well-suited for practical scenarios such as the Web of Things (WoT). However, to effectively identify malicious gradients, most existing secure schemes that defend against poisoning attacks rely on additional data assumptions (e.g., a small clean dataset on the server or benign users with clean data), which are often impractical in real-world applications. In this paper, we propose SecHFed, a secure framework that achieves gradient security and resists poisoning attacks without data assumptions in HFL. The core idea of SecHFed is to amplify gradient differences to mitigate the influence of anomalous gradients within the ciphertext domain. Specifically, we employ Fully Homomorphic Encryption (FHE) to encrypt gradients for a good balance between security and efficiency. In the HFL setting, we utilize two non-colluding edge nodes to verify and process encrypted gradients without exposing sensitive information to either node. After calculating the gradient weights based on the similarity scores and the Softmax function, the nodes and the cloud server aggregate encrypted gradients. We formally provide mathematical proofs and analyses of convergence, perturbation, security, and complexity of SecHFed. Finally, we evaluate SecHFed on three real-world datasets. Extensive experimental results demonstrate the effectiveness of SecHFed in defending against various poisoning attacks.","PeriodicalId":50367,"journal":{"name":"Information Fusion","volume":"9 1","pages":""},"PeriodicalIF":18.6,"publicationDate":"2026-08-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148884147","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":"PRSL: Privacy-Preserving and Robust Split Learning for Multi-Source Embedding Fusion","authors":"Mingtian Jia, Yiran Li, Caixu Xu, Yanping Wang, Jie Zhou, Shengke Zeng","doi":"10.1016/j.inffus.2026.104733","DOIUrl":"https://doi.org/10.1016/j.inffus.2026.104733","url":null,"abstract":"For privacy-sensitive collaborative learning with vertically partitioned data, split learning has attracted increasing attention as a multi-source embedding fusion paradigm. In this setting, distributed clients generate local embeddings from distinct feature subspaces, while the server fuses them for prediction and training. However, under privacy constraints and adversarial conditions, this fusion paradigm faces two key challenges: exposed plaintext embeddings may leak sensitive information from clients, and poisoned local embeddings may compromise the reliability of the fused representation. Existing split learning studies typically address privacy preservation and robustness in isolation. However, jointly preserving the confidentiality of local embeddings and achieve poisoning-resistant fusion remains less explored. To address this problem, we propose PRSL, a privacy-preserving and robust split learning framework for multi-source embedding fusion. Specifically, PRSL establishes a privacy-preserving framework based on non-interactive efficient decryption functional encryption to preserve the confidentiality of local embeddings. Within this protected framework, PRSL further designs a reliability-aware robust aggregation mechanism to mitigate the impact of poisoned local embeddings on multi-source fusion. Security analysis demonstrates that PRSL prevents the recovery of any individual plaintext local embedding beyond authorized functional outputs. Experiments on Bank and MNIST show that PRSL improves model accuracy by 30.20%–80.97% over competing methods under untargeted attacks and reduces the attack success rate to as low as 7.69% under targeted attacks, while maintaining stable performance in both IID and non-IID settings. PRSL also demonstrates lower computation and communication overhead than the evaluated alternatives based on homomorphic encryption.","PeriodicalId":50367,"journal":{"name":"Information Fusion","volume":"32 1","pages":""},"PeriodicalIF":18.6,"publicationDate":"2026-08-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148853345","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}