15 October 2026, Volume 52 Issue 10
-
Select all|
-
Computer Engineering. 2026, 52(10): 1-17. https://doi.org/10.19678/j.issn.1000-3428.0260693
The smart grid is a critical infrastructure in the modern energy system that couples cyberspace with physical operation processes, and it has become a key pillar of national energy security. With the widespread use of massive heterogeneous devices such as distributed energy resources, energy storage systems, and smart meters, the "cloud—network—edge—terminal" collaborative architecture has become the mainstream paradigm. However, the continuous evolution of attack techniques has led to increasingly frequent cyberattacks targeting Supervisory Control and Data Acquisition (SCADA) systems, intelligent terminals, and communication networks. The inherent lag of conventional ″patch-and-block″ passive defense techniques and their dependence on prior knowledge cause difficulty in effectively countering security threats deeply embedded in system architectures and operational processes. Endogenous security issues caused by inherent defects such as vulnerabilities and backdoors in information systems cannot easily be completely eliminated during the design and development phases; this creates a fundamental bottleneck that restricts the improvement of smart grid security protection capabilities. To address these issues, this paper systematically reviews endogenous security technologies that can be deployed in smart grids. First, the six major threats facing smart grids—supply-chain attacks, vulnerability exploit attacks, Distributed Denial of Service/Denial of Service (DDoS/DoS) attacks, data tampering and forgery attacks, physical and side-channel attacks, and Advanced Persistent Threats (APTs)—are considered as research entry points. Typical entry points, impact levels, and the technical difficulties and limitations of traditional defenses for each type of attack are analyzed, combined with the Smart Grid Architecture Model (SGAM) and typical attack cases. The analysis reveals that the direct sources of most attacks are concentrated at the component and communication layers, and that the attack impacts can propagate across layers. Second, according to core functional positioning, deployable endogenous security technologies are categorized into attack-resistant and trusted authentication technologies. Attack-resistant technologies include software diversification, Dynamic Heterogeneous Redundancy (DHR), Physical Unclonable Function (PUF)-based authentication mechanisms, and attack-resistant routing, while trusted authentication technologies include zero-trust architecture, trusted computing, and quantum-secure trusted authentication. This paper systematically reviews the research progress, defense objectives, operational layers, and verification results of these seven technologies based on their working mechanism, applicable threats, deployment layers, performance overhead, and engineering conditions. It also summarizes the applicable conditions and main limitations of each technology. The review reveals five key findings. First, a single endogenous security technology usually acts on only partial segments of an attack chain and cannot independently cope with composite attacks such as APTs. The key to multi-technology collaboration lies in clarifying the division of labor according to attack phases and the sharing of information such as identities, device statuses, abnormal outputs, and network statuses. Follow-up research should evaluate collaboration effectiveness using metrics such as attack chain blocking rate, false judgment rate, worst-case latency, and recovery time. Second, most existing technical solutions are still in the prototyping, simulation, or small-scale verification stages, and large-scale deployment verification has not been performed in real smart grid scenarios, where performance overhead and real-time constraints are key obstacles to technology implementation. Third, supply chain security is the foundation for the reliable operation of endogenous security technologies. Software diversification relies on a trusted compilation chain, while dynamic heterogeneous redundancy relies on trusted execution body sources, and zero-trust and trusted computing solutions rely on accurate asset and firmware information. Accordingly, a closed-loop process from component identification and integrity verification to anomaly response should be established. Fourth, Artificial Intelligence (AI) can assist endogenous security detection and authentication, but both recognition capability and model security must be considered. Evaluations should simultaneously report detection performance, inference latency, resource overhead, and update mechanisms. Fifth, quantum security technologies should be migrated progressively according to their current development status. Post-Quantum Cryptography (PQC) can be prioritized for deployment via pure software solutions with cryptographic agility mechanisms, while Quantum Key Distribution (QKD) can be used for priority application in fixed high-value links.
-
Computer Engineering. 2026, 52(10): 18-30. https://doi.org/10.19678/j.issn.1000-3428.0252595
The Log-Structured Merge-tree (LSM-tree) is widely employed in key—value storage systems, leveraging its sequential write mechanism to achieve efficient write performance. However, it also introduces challenges such as high read—write amplification, significant overhead from merge operations, and data redundancy. Traditional optimization approaches enhance the system performance by adjusting tree structures, refining merge strategies, and adopting key—value separation mechanisms. Nevertheless, in the era of big data, with the explosive growth in data volume, LSM-tree must handle more frequent write and merge operations, leading to a sustained strain on CPU computing resources, which gradually becomes a bottleneck for system performance improvement. Furthermore, traditional optimization methods fail to eliminate the substantial I/O operations between the host and storage devices and still face high costs associated with redundant data migration. Computational storage technology offers new perspectives for addressing these challenges by deploying additional computational resources at the storage layer, alleviating CPU burdens through task offloading and further reducing performance degradation caused by data migration through near-data processing. This study focuses on the optimization of an LSM-tree based on computational storage technology. First, it outlines the architecture of computational storage technology. Subsequently, it categorizes and compares existing solutions from two perspectives—merge operation optimization and data migration optimization—to address the primary bottlenecks faced by systems in the context of big data. Finally, considering the limitations and developmental trends of the current research, this study provides insights into future research directions.
-
Computer Engineering. 2026, 52(10): 31-49. https://doi.org/10.19678/j.issn.1000-3428.0253234
In recent years, generative image technology based on Diffusion Models (DMs) has made breakthroughs, with text-to-image models represented by Stable Diffusion, DALL-E, and Midjourney have been widely applied in commercial sectors and everyday life. However, highly realistic artificial intelligence-generated images have also brought challenges to image authenticity, giving rise to social issues such as the spread of misinformation and copyright infringement. To address these challenges effectively, this paper systematically reviews the latest research progress in detection technologies for images generated by DMs. First, it outlines the development trajectory of DMs based on their principles and fundamental frameworks for large-scale applications. Second, it summarizes the evolution of dataset construction, indicating that dataset development is progressing from using few generators and low resolutions to multi-model integration and high-quality, multi-stage filtering. Third, it analyzes three types of detection technologies: those based on implicit, explicit, and hybrid features. Finally, it analyzes the main challenges facing current detection technologies and provides an outlook on future research directions. This review offers researchers and practitioners with a comprehensive technical landscape and reference for future development trends.
-
Computer Engineering. 2026, 52(10): 50-72. https://doi.org/10.19678/j.issn.1000-3428.0253044
Short-term photovoltaic power forecasting is the foundation for optimizing power dispatch in power systems, and its accuracy directly affects the overall system efficiency. However, photovoltaic power, which is affected by multiple meteorological factors, exhibits short-term volatility and randomness, posing severe challenges for high-precision forecasting. In recent years, Deep Learning (DL) has shown an excellent ability to excavate intrinsically correlated features, offering a promising technical path for precise forecasting. The application of DL to short-term photovoltaic power forecasting was studied systematically. First, a typical application paradigm for mainstream deep learning models is presented. Subsequently, the application status and performance of deep neural networks, such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Transformers, and Graph Neural Networks (GNNs), in the task are discussed considering the ideal scenario of sufficient available data with static and no incremental changes. Furthermore, the strategies and progress of deep neural network based on technologies such as deep data augmentation, Transfer Learning (TL), Federated Learning (FL), and Online Learning (OL) are analyzed to address real-world challenges such as data scarcity and limited access. Finally, the challenges regarding robustness, generalization, and adaptation faced by existing research are discussed along with future research routes from the perspectives of forecasting architecture and optimization strategies.
-
Computer Engineering. 2026, 52(10): 73-80. https://doi.org/10.19678/j.issn.1000-3428.0252115
Combinatorial optimization problems have important applications in areas such as logistics path planning; however, their solution space exponentially expands with the problem size, leading to severe challenges for traditional methods. In recent years, neural combinatorial optimization methods based on reinforcement learning have been able to achieve solution quality close to that of traditional solvers while keeping the solution computation time short. The mainstream method, called Policy Optimization with Multiple Optima (POMO), enhances the training stability through symmetry optimization; however, its unidirectional sequence generation mechanism still suffers from two limitations: the traditional constructive method cannot fully exploit the symmetry features of the problem easily and the endpoint information cannot effectively participate in the decision-making process of the remote node. To address this problem, this paper proposes a Bidirectional Construction Strategy (BCS)-based POMO model, named BCS-POMO, which dynamically selects the extension direction with higher confidence by constructing the solution sequence in parallel from the start and end points, avoiding models that are caught in a dilemma owing to unidirectional constructions. The model exploits the symmetry of the construction sequence to achieve weight parameter sharing and improves efficiency through batch parallel computation. Experimental results show that the BCS-POMO effectively reinforces the role of endpoint information as a decision aid in the construction process, which reduces the error by 16% and 18% for the Traveling Salesman Problem (TSP) and the Capacitated Vehicle Routing Problem (CVRP), respectively, verifying the effectiveness of the bidirectional construction strategy in exploiting endpoint information and the advantages of symmetry modelling.
-
Computer Engineering. 2026, 52(10): 81-93. https://doi.org/10.19678/j.issn.1000-3428.0070696
Knowledge graphs, a graph-based data organization method, provide rich semantic information and contextual background for recommendation systems, enabling them to effectively handle complex user behaviors and item features. Existing knowledge graph-based recommendation methods still face challenges such as information over-smoothing and handling of anomalies, particularly in large-scale data processing scenarios. Over-smoothing often prevents the model from capturing personalized user needs, and interference from anomalies can affect the accuracy and robustness of the recommendations. To address these issues, this paper proposes a knowledge graph recommendation model based on user behavior feature fusion and anomaly detection. The model effectively avoids information over-smoothing by incorporating fused user behavior features. Additionally, the model integrates an anomaly detection mechanism that identifies and removes noisy data and abnormal behaviors, significantly improving the accuracy and robustness of the recommendations while reducing the impact of poor data on the recommendation results. To validate the effectiveness of the model, experiments are conducted on three real-world datasets. The experimental results show that, compared to the existing best baseline models, the proposed model achieves an average improvement of 0.067 7 in Area Under the Receiver Operating Characteristic Curve (AUC) and 0.050 9 in F1 values. The performance improvement is particularly significant in datasets with high sparsity, effectively alleviating the challenges posed by data sparsity.
-
Computer Engineering. 2026, 52(10): 94-105. https://doi.org/10.19678/j.issn.1000-3428.0252002
Conversational Recommender System (CRS) can provide personalized recommendations through multi-round user interactions. It has broad application prospects in fields such as e-commerce and movie consultation. However, existing studies neither specifically extract the association information between items in actual dialogues nor consider the weight distribution when words and entities coexist. This results in the inability of the system to comprehensively extract the semantic information reflected in the conversation text and consequently affects its ability to generate fluent responses and personalized recommendations. To address these limitations, this study proposes Semantic Enhanced Conversational Recommender System (SECRS), which tunes the pretrained language model DialoGPT with tunable prompts to perform conversational recommendation tasks. First, this study develops a session-enhanced item extraction method using graph neural networks to enhance the semantic information of item-level entities, which mines the transitivity and co-occurrence relations among items. Second, it introduces an adaptive semantic fusion method, which weighs session-enhanced item embeddings and entity embeddings, then fuses the entity embeddings before and after session enhancement with the word embeddings by independent mapping and joint fusion, respectively, to constitute the semantic-enhanced prompts in the conversational task and recommendation task. These methods improve the linguistic expressions and personalized recommendation capabilities of the system. The experimental results on two datasets show that SECRS outperforms the mainstream CRSs. Compared with the baseline model UniCRS, SECRS achieves an 8.33% increase in Recall@10 and a 45.15% improvement in Distinct-4 on the ReDIAL dataset. Comparison and ablation experiments demonstrate that the SECRS can generate responses with more satisfying recommendations and a more diverse language.
-
Computer Engineering. 2026, 52(10): 106-115. https://doi.org/10.19678/j.issn.1000-3428.0070397
Aspect-Based Sentiment Analysis (ABSA) aims to determine the sentiment polarity of specific aspect terms in sentences with complex syntactic structures. Existing models based on dependency trees and Graph Neural Networks (GNN) suffer from difficulty in fully extracting syntactic structures and deep semantic features, and their feature fusion mechanisms fail to effectively integrate semantic features with syntactic structures, resulting in low accuracy in sentence-level sentiment polarity judgment. To address this issue, a novel ABSA model based on DeBERTa is proposed. First, DeBERTa is employed to generate textual word embeddings, whereas an aspect-aware attention mechanism is utilized to extract aspect term features and abstract semantic representations are leveraged to capture textual syntactic structures, thereby reducing the impact of incomplete feature extraction on subsequent sentiment analysis. Second, a novel triangular multiplicative fusion mechanism that integrates syntactic structures with deep semantic features is constructed. Finally, through a triangular self-attention mechanism and fully connected networks, the sentiment polarity features of aspect terms are mapped to a sentiment classification layer, effectively suppressing the interference of irrelevant noise and thereby improving the accuracy of the sentiment polarity judgment. Experimental results demonstrate that, compared with the latest baseline models, the proposed model achieves an average improvement of 0.93 percentage points in sentiment polarity judgment accuracy and 1.39 percentage points in the macro-averaged F1 value. This model can effectively capture both the syntactic structures and deep semantics of sentences, yielding high accuracy in sentiment polarity classification.
-
Computer Engineering. 2026, 52(10): 116-128. https://doi.org/10.19678/j.issn.1000-3428.0070761
Unsupervised Domain Adaptation (UDA) aims to transfer knowledge from a labeled source domain to an unlabeled target domain, thereby improving the performance of the target domain model. However, traditional UDA methods assume that the category spaces of the source and target domains are identical; hence, they cannot handle unknown categories present in the target domain, which limits their applicability in real-world scenarios. Open-Set Domain Adaptation (OSDA) addresses this issue by introducing the recognition of unknown categories. Nevertheless, effectively mitigating the negative impact of inter-domain discrepancies and class imbalances on the model performance remains a significant challenge. Existing OSDA methods tend to overlook domain-specific features and directly minimize domain discrepancies, which may lead to ambiguous decision boundaries between classes and weaken the generalization ability of the model. To address this problem, this paper proposes an open-set domain adaptation method based on Optimal Transport Distance Regularization (OTR) and neighborhood clustering, called OTRNC. Specifically, the method employs OTR to maximize the distributional distance between high-confidence and low-confidence sample groups, thereby reducing the interference of unknown categories in the domain adaptation process. Subsequently, it leverages dynamic neighbor retrieval and invariant feature learning to reduce intra-class variations in the target domain and enhance the generalization capability of the learned features. Experimental results demonstrate that OTRNC achieves superior performance across multiple benchmark datasets, thereby validating its effectiveness.
-
Computer Engineering. 2026, 52(10): 129-137. https://doi.org/10.19678/j.issn.1000-3428.0252004
To address the privacy leakage risks associated with collecting and annotating sensor data for wearable device-based fitness action recognition as well as the limitations of traditional centralized model training in adapting to new users, this study proposes a fitness action recognition method based on a wearable sports knee sleeve and personalized Federated Learning (FL). The method accomplishes two tasks, namely fitness action type recognition and user performance level recognition, to help users understand their fitness performance and improve exercise effectiveness. First, each user is treated as an independent task, and a global embedding network is meta-trained in a federated manner to learn user-agnostic shared representations that can be effectively generalized for any user. Subsequently, through an adaptation procedure, the local classification network is fine-tuned in two stages on top of the global embedding network, enabling each user to obtain a personalized model. Finally, extensive experiments are conducted on a real-world fitness dataset to validate the performance of the proposed method. The experimental results demonstrate that the proposed method achieves 100% accuracy in fitness action type recognition and 95.94% accuracy in user performance level recognition, significantly outperforming existing state-of-the-art methods. The proposed system can protect user privacy while exhibiting a strong generalization ability for new users.
-
Computer Engineering. 2026, 52(10): 138-153. https://doi.org/10.19678/j.issn.1000-3428.0070676
The accurate generation of Multivariate Time Series (MTS) data provides an effective way to solve the problem of insufficient data scale and is crucial for downstream tasks such as power load forecasting and wind and solar power generation evaluation. However, existing methods cannot capture the long- and short-term dependence and correlation between variables easily, and they cannot meet the requirements of energy system analysis owing to a lack of interpretability. Therefore, an interpretable diffusion model for multivariate time series generation method IDMTS is proposed. First, a transformer architecture containing Triplet Attention (TA) is introduced into the denoising network of the diffusion model to capture long- and short-term dependencies and variable feature associations. Subsequently, combined with multiscale trend seasonal decomposition, the trend and season terms are modeled using Bidirectional Long Short-Term Memory (BiLSTM) network and Fourier Attention (FA), respectively. This improves the accuracy and interpretability of generated data. Simultaneously, the generation quality is optimized by multiscale Adaptive Maximum Mean Discrepancy (Ada-MMD) loss function. The experimental results show that the generation accuracy of IDMTS on four public datasets is significantly better than that of the baseline method, in which the Context-FID score, correlation score, discriminant score, and prediction score are reduced by 51.5% to 84.5%, 4.1% to 26.8%, 24.0% to 68.8%, and 0.3% to 40.0%, respectively. Additionally, IDMTS shows good interpretability and generalization ability in the interpretability, conditional interpolation, and prediction experiments.
-
Computer Engineering. 2026, 52(10): 154-163. https://doi.org/10.19678/j.issn.1000-3428.0070675
Chinese medical Named Entity Recognition (NER) aims to identify entities with specific meanings from medical texts, such as diseases, drugs, symptoms, and anatomical parts. This task provides robust support for clinical decision making, medical information integration, and medical record management. However, existing research on Chinese medical NER has not fully addressed the complexity of medical texts, which are characterized by an abundance of specialized terminology, limited embedding diversity, and insufficient utilization of semantic information. To address these issues, this paper proposes a Chinese medical NER model that integrates multigranularity features. The model first employs a BERT pretrained model to generate character embeddings for the text. It then uses both one-dimensional and two-dimensional convolutional neural networks to extract character shape and stroke features, while external lexicons are incorporated to introduce word-level features, enhancing the representation of word and entity boundaries. Additionally, sentence-level features are included to capture global semantic information. A cross-attention mechanism is utilized to iteratively fuse these multigranular features, resulting in embeddings enriched with deep semantic information. Finally, Conditional Random Fields (CRF) are used to output the entity recognition results. Experimental results on the CCKS2017 and CCKS2019 datasets demonstrate that the proposed model achieves F1 values of 92.88% and 87.86%, respectively, outperforming mainstream models in recognition performance.
-
Computer Engineering. 2026, 52(10): 164-176. https://doi.org/10.19678/j.issn.1000-3428.0252596
Object detection is a fundamental task in computer vision, which aims to locate and classify objects within an image. By contrast, multi-attribute recognition involves inputting an image centered around an object and outputting the classification results of multiple attributes of the object within the image. Current detection algorithms face challenges in recognizing small objects, while attribute recognition algorithms fail to locate objects. To address these issues, this paper proposes a multi-attribute small object detection framework, MA-YOLO, that performs small object detection and multi-attribute recognition simultaneously. First, an Attribute Decoupling Module (ADM) decouples attributes of the object and adds multiple parallel attribute classification heads. Second, a simple yet effective attribute-fusion strategy is used to facilitate feature interactions among the attributes. This strategy generates bridging features through a Bridge Feature Generation Module (BFGM) to bridge the gap between the features extracted by the backbone network and all decoupled attribute features, and it utilizes an Attribute Fusion Module (AFM) to obtain fused attribute features. For small object detection, a convolution and self-attention hybrid module, ACmix, is adopted to enhance the feature extraction capability of the network. Additionally, bounding boxes are modeled as two-dimensional Gaussian distributions and measured using the Normalized Wasserstein Distance (NWD). Experimental results show that MA-YOLO achieves mAP@0.5 values of 84.1% and 85.3% on the test and validation sets of the multi-attribute dataset WIDER Attribute and 48.2% as the validation set of the small object dataset VisDrone2019. Extensive ablation and visualization experiments validate the effectiveness of MA-YOLO.
-
Computer Engineering. 2026, 52(10): 177-187. https://doi.org/10.19678/j.issn.1000-3428.0252645
Sketch-guided image inpainting has significant application value in photo restoration and creative editing. However, it faces the dual challenges of scarce user sketch data and restoration distortions caused by geometric deviations. Existing methods rely on edge detection to generate pseudo-sketches while neglecting user-drawn deviations (e.g., hand tremors and stroke breaks), leading to structural misalignment and detail blurring in complex scenes. To address these challenges, this study proposes an innovative framework that combines a Deformable Sketch generation Network (DSN) with dual-stage-guided inpainting. First, a DSN is constructed to model typical hand-drawn deviations, generating a large-scale sketch-image paired dataset with realistic geometric deformation features and effectively alleviating data scarcity. Second, a two-stage inpainting framework is designed: the first stage corrects geometric misalignment and repairs structural breaks in input sketches to optimize the sketches, whereas the second stage effectively integrates the optimized sketch information into the inpainting network to achieve collaborative optimization of global structural constraints and local texture generation. Experiments on benchmark datasets validate the effectiveness of the method, achieving a Peak Signal-to-Noise Ratio (PSNR) of 25.78 dB and Structural Similarity Index Metric (SSIM) of 0.852 on the CelebA-HQ dataset. The results demonstrate that this method effectively addresses the challenges of scarce user sketch data and geometric deviations while significantly improving the structural accuracy and perceptual quality of sketch-guided image inpainting.
-
Computer Engineering. 2026, 52(10): 188-198. https://doi.org/10.19678/j.issn.1000-3428.0252270
Learning disentangled representations to enhance the controllability of image generation models is a key research direction in computer vision. However, existing methods have two major limitations: reliance on large-scale annotated data and difficulty in handling complex dependencies between features. To address these issues, this study proposes a universal generative disentangling method based on the Hilbert—Schmidt Independence Criterion (HSIC). This method innovatively converts HSIC into an independence regularization mechanism for the latent space of generative models. By incorporating HSIC regularization terms, it optimizes the measurement objective of nonlinear dependency relationships and guides the model in learning independent feature representations. Specifically, the method integrates HSIC into two mainstream generative model architectures. For the Variational AutoEncoder (VAE) class, it combines variational inference with HSIC regularization to optimize latent distribution disentanglement. For the Diffusion Model (DM) class, it gradually achieves progressive feature disentangling by embedding the HSIC regularization term into the time-step optimization of the reverse process. The experimental results show that this universal method, which can be implemented in different model architectures, enhances the latent representation independence and maintains stable performance in unsupervised settings, offering a novel approach to model complex feature dependencies. To further verify the semantic consistency of the disentangling space, latent space interpolation experiments are conducted to generate smoother trajectories, demonstrating that HSIC regularization constructs a linearly separable disentangling space. Additionally, a dual validation using standard disentangling metrics and HSIC-based custom metrics is performed, which reveals a positive correlation and confirms the objectivity of the disentangling evaluation criteria.
-
Computer Engineering. 2026, 52(10): 199-214. https://doi.org/10.19678/j.issn.1000-3428.0252062
Because of the prolonged examination time and limited therapeutic time window for stroke, the development of a rapid and highly accurate medical image segmentation model for stroke is important for clinical diagnosis. The U-Net architecture based on Mamba, which is known for its low complexity and capability to handle large-scale images, has garnered widespread attention in the field of medical image processing in recent years. The FRactional order Fourier Transform (FRFT) can convert signals to arbitrary fractional domains between the spatial and frequency domains, allowing the observation of features that are not prominent in the spatial or frequency domains. Therefore, by introducing the FRFT, the lesion characteristics can be observed in the fractional domain. Based on the FRFT and Mamba network, a novel model named FRFTMamba-UNet is proposed for stroke medical image segmentation. This model incorporates a fractional domain into the Mamba network and designs a multilevel residual module connected to a U-Net encoder. In addition, a hierarchical feature extraction strategy is implemented in a U-Net-like network, where different feature extraction modules are designed for shallow and deep layers. Specifically, residual convolutions based on Convolutional Neural Networks (CNN) are added to the shallow layers to extract shallow features effectively, whereas the Mamba architecture is utilized in the deep layers to extract deep features further. The proposed method demonstrates superior accuracy and efficiency to existing state-of-the-art models based on the Mamba module across three stroke datasets: AISD, ATLAS, and ISLES 2022. On the AISD dataset, it achieves a Dice Similarity Coefficient (DSC) of 64.27%, 62.24%, whereas on the ISLES 2022 dataset, it achieves a DSC of 85.24%.
-
Computer Engineering. 2026, 52(10): 215-226. https://doi.org/10.19678/j.issn.1000-3428.0252028
The measurement of physiological parameters using facial videos in online education is currently a research hotspot in intelligent education. Traditional remote Photoplethysmography (rPPG) cannot adapt to changes in the lighting environment in online educational scenarios, which affects the flexibility and accuracy of physiological parameter measurements based on facial videos. Considering typical lighting scenarios in online education, a method for extracting Blood Volume Pulse (BVP) signals based on lighting adaptability is proposed, and a dual correction model for BVP signals is constructed by combining a Generative Adversarial Network (GAN) and a Convolutional Neural Network (CNN). First, the optimal solution for the orthogonal chrominance signal under different lighting conditions is calculated based on the simulated annealing algorithm. Simultaneously, a lighting scene prediction mechanism for classifying lighting scenes using the average gray level intensity is established to achieve the optimal chrominance signal that adapts to the lighting scene. Furthermore, the GAN and CNN models are combined to perform dual correction on the BVP signal to ensure that the final output physiological parameters are accurate and reliable. The model is verified using four publicly available datasets reorganized for typical educational scenarios. The experimental results show that the Root Mean Square Error (RMSE) of Heart Rate (HR) is reduced by an average of 8.3 bpm, demonstrating the robustness and accuracy of the model under different lighting conditions. This model has significant advantages in terms of improving the accuracy of heart rate and heart rate variability prediction and can provide effective support for contactless physiological parameter detection in complex lighting environments.
-
Computer Engineering. 2026, 52(10): 227-241. https://doi.org/10.19678/j.issn.1000-3428.0252123
The human visual cortex has a hierarchical structure in which binocular fusion and binocular rivalry first occur in low-level visual areas. However, current deep learning-based Stereoscopic Image Quality Assessment (SIQA) models generally estimate the quality values of stereoscopic images by fusing features of left- and right-view images at different levels of the network, resulting in insufficient simulation of the perceptual mechanisms in the low-level visual areas of the human visual cortex. To address this issue, this study proposes a stereoscopic image quality assessment method that simulates binocular rivalry. First, to simulate binocular visual rivalry, a binocular image fusion model based on an unsupervised approach is constructed by leveraging the ability of deep Convolutional Neural Networks (CNNs) to acquire prior knowledge of the input images. Considering the left and right views as learning targets, this model is built on an encoder—decoder architecture, and it simulates the binocular fusion process in the human visual system. The gradient magnitude responses of the left and right images are used to evaluate the degree of image degradation and to determine the fusion weights of the left and right views, which simulate the binocular rivalry phenomenon. Second, a ResNet50 model pretrained on a large-scale image database is used to extract quality-aware features from the fused image and construct a feature-quality mapping model based on Support Vector Regression (SVR) for estimating the quality value of the stereoscopic image. Experimental results on four classic stereoscopic image benchmark databases show that the proposed method achieves over 0.96 in both the Pearson Linear Correlation Coefficient (PLCC) and Spearman Rank Order Correlation Coefficient (SROCC), and its Root Mean Square Error (RMSE) is superior to that of the compared methods. These results indicate that the proposed unsupervised binocular image fusion method can effectively mimic binocular visual effects, thus significantly improving the accuracy of stereoscopic image quality assessment.
-
Computer Engineering. 2026, 52(10): 242-251. https://doi.org/10.19678/j.issn.1000-3428.0252150
In Image—Text person Re-IDentification (ITReID) tasks, initializing models with parameters from pretraining models has become a mainstream paradigm that effectively alleviates the feature alignment bottleneck of single-modal models caused by a lack of cross-modal information. Existing methods focus on mining semantic features at different scales in an image-text joint embedding space for optimization. However, the introduction of the new alignment paradigm is prone to causing the pretraining model to fall to a local minimum during fine-tuning. To address these issues, this paper proposes a Prompt-based Information Transfer (PIT) framework. By introducing cross-modal prompt tokens into the original forward process of the single-modal encoder and cross-modal image-text encoder, the proposed framework promotes early feature fusion and implicitly guides the pretraining model to focus more on modal-invariant information. PIT includes a prompt-based Contrastive Learning (CL) loss and Prompt Training Strategy (PTS). Prompt-based contrastive loss aims to construct a shared feature-embedding space with both intra-modal discrimination and inter-modal semantic consistency by constraining the similarity between graphic and text features. The PTS can be regarded as a form of self-distillation that treats the pseudo-targets generated by nonprompt features and ground truth as another view of image-text pairs, supervises the training process, and makes the learned embeddings contain richer multimodal information. Only 0.61×106 additional parameters are introduced based on fine-tuning, and PIT achieves Rank-1 improvements of 1.48, 1.5, and 1.55 percentage points on three public datasets.
-
Computer Engineering. 2026, 52(10): 252-262. https://doi.org/10.19678/j.issn.1000-3428.0252234
In Fine-Grained Image Classification (FGIC) tasks, a sufficient number of samples can provide rich local feature information. However, in few-shot scenarios, data sparsity makes it difficult for models to fully capture discriminative local information. To address this problem, a Few-Shot Learning (FSL) method integrating axial attention and scale-aware mechanisms is proposed. First, a Frequency-Adaptive Feature Selection (FAFS) module is designed to reduce background noise and interference from nontarget regions, highlight discriminative local features, and thereby increase feature separability among different categories. Second, an axial-scale joint enhancement module is constructed to fuse global contextual information, focus on key regions, process features with different receptive fields in parallel, and enhance the representation capability of the model for details at various scales. Finally, a dual-similarity metric module is adopted, which guides learning through two similarity metric methods to improve feature generalization and reduce bias toward specific features. On public datasets CUB and Stanford Dog, compared with the baseline method, the proposed method improves the classification accuracy by 1.4 and 1.45 percentage points under 1-shot and 5-shot settings, respectively, and by 1.86 and 3.49 percentage points on the other dataset. On the Stanford Car dataset, the proposed method achieves the best performance under the 1-shot setting and obtains competitive results under the 5-shot setting. Experimental results demonstrate that the proposed method effectively improves the performance of Few-Shot Fine-Grained Image Classification (FSFGIC) and is capable of better capturing discriminative feature information.
-
Computer Engineering. 2026, 52(10): 263-274. https://doi.org/10.19678/j.issn.1000-3428.0252029
Object detection and multi-target tracking technologies are becoming increasingly mature. However, when performing aerial multi-target tracking tasks in complex scenarios, issues such as small target size, large size variation, and occlusion lead to unsatisfactory detection and tracking performances. Therefore, this paper proposes an aerial multi-target tracking algorithm based on an improved YOLOv8 and ByteTrack method called YBTrack. First, a detector called Multi-Scale Attention YOLO (MSA-YOLO) is constructed. The original convolution in YOLOv8 is replaced with a space-depth convolution that transforms spatial information into channel dimensions, effectively preserves target details, and reduces missed and false detections caused by information loss during multiscale feature map fusion. Simultaneously, a lightweight accelerated space-channel attention module is designed for neck convolution to reduce computational complexity. This module also acts as a feature refinement module before the detection head, further enhancing its ability to extract target feature information. Next, the ByteTrack tracking model is optimized to improve the tracking performance. A Spatial-Appearance Similarity Matrix (ASM) is designed to enhance the model's ability to distinguish similar targets. In addition, a target correction function is proposed to reduce the error accumulation of the Kalman filter and decrease the target offset and loss rates. Finally, MSA-YOLO and the optimized ByteTrack are combined for multi-target tracking experiments. MSA-YOLO achieves a 9.4 percentage points improvement in mAP@0.5 on the VisDrone2019-DET dataset. The multi-target tracking algorithm improves the Multiple Object Tracking Accuracy (MOTA) by 11.2 and 8.3 percentage points and the Identification F1 Score (IDF1) by 8.9 and 7.4 percentage points on the VisDrone2019-MOT and MOT17 datasets, respectively. Experimental results demonstrate the excellent tracking performance of the proposed method. Furthermore, comparison experiments with other multitarget tracking algorithms confirm the superiority of the proposed algorithm.
-
Computer Engineering. 2026, 52(10): 275-284. https://doi.org/10.19678/j.issn.1000-3428.0252352
Multimodal sentiment analysis utilizes multimodal data to infer human emotions. However, existing models significantly degrade in performance when faced with issues such as modal information loss, text dependency, and cross-modal conflicts. To address this, a Generative Completion and Dynamic Knowledge Fusion Model (GC-DKF) for multimodal sentiment analysis is proposed. First, the model employs a generative prompt learning module to complete the missing intra-modal and inter-modal information in the original data, generating missing modal features to enhance the model's adaptability to uncertain modal scenarios. Then, a dynamic dominant modality selection mechanism is designed to dynamically select the dominant modality based on emotional proportion factors. Meanwhile, a knowledge encoder is introduced to strengthen the representation capability of a single modality, obtaining knowledge-enhanced representations of each modality. Finally, guided by the features of the dominant modality, the model further learns other secondary modalities to generate complementary multimodal joint representations, achieving more efficient and accurate multimodal sentiment analysis. Experiments on the public CMU-MOSI and CMU-MOSEI datasets demonstrate that the proposed model outperforms existing mainstream multimodal sentiment recognition methods in terms of metrics such as binary classification accuracy, F1-score, Mean Absolute Error (MAE), and Pearson correlation coefficient, with sentiment recognition accuracies reaching as high as 83.55% and 83.02%, respectively. This fully demonstrates that the proposed model has strong competitiveness in multimodal sentiment recognition tasks.
-
Computer Engineering. 2026, 52(10): 285-301. https://doi.org/10.19678/j.issn.1000-3428.0252046
With the rapid development of Internet technology, personalized recommendation systems have become pivotal in helping users filter content of interest. Traditional recommendation methods have limitations in processing large-scale data and capturing the complex preferences of users. Existing recommendation methods based on Graph Neural Networks (GNNs) primarily focus on mining direct interactions between users and items. Although they improve the accuracy of recommendations, they often ignore the integration and utilization of multimodal information such as text, images, audio, and videos. Metapath, which is a concept that describes the composite relationship between nodes in heterogeneous graphs, can further improve the embedding quality and recommendation effect. However, the existing models either ignore node content features, discard intermediate nodes on the metapath, or consider only a single metapath. To address the challenges of existing multimodal recommendation systems, this study proposes a multimodal recommendation algorithm based on metapath guidance MAMGNN. First, it constructs a multimodal heterogeneous information network to integrate information from different modalities and then uses metapaths to guide the propagation and aggregation of information within the intra-metapath and inter-metapath. Furthermore, it introduces GNNs and attention mechanisms to learn high-quality embedding representations of users and items, generating more accurate and explainable recommendation results. Extensive experiments on two real-world datasets, namely, MovieLens-20M and H&M, and a small-scale user survey demonstrate that the MAMGNN significantly enhances the performance in predicting user preferences for items, outperforming suboptimal models in the Precision@10, Recall@10, and NDCG@10 metrics by approximately 2.93%, 1.98%, 2.12%, and 3.43%, 1.18%, and 2.40%, respectively.
-
Computer Engineering. 2026, 52(10): 302-316. https://doi.org/10.19678/j.issn.1000-3428.0260521
In real-world clinical consultations, patients' chief complaints are typically expressed verbally and subsequently recorded as text by physicians. Physicians must comprehensively use both patients' spoken descriptions and the corresponding textual records to judge and classify symptoms, thereby providing a basis for subsequent clinical decision making. However, this method has several limitations. Speech information is susceptible to environmental noise and individual pronunciation differences, whereas textual records are unable to fully reflect speech-related expressive features such as speaking rate, pauses, and intonation. Additionally, patients' chief complaints are usually colloquial, subjective, and unstructured, and the semantic boundaries among different symptom categories may be ambiguous. Therefore, single-modality methods cannot achieve satisfactory classification performance easily. To address these issues, a dynamic weight decision fusion-based multimodal symptom classification method, called DWDF-MSC, is proposed to fully exploit the complementarity between textual and speech information and improve the accuracy and robustness of symptom classification. The proposed method consists of three stages: multimodal feature extraction, preliminary classification, and adaptive gated decision fusion. In the multimodal feature extraction stage, text and speech branches are constructed to model the text and speech data of the patient's chief complaint in parallel. In the text branch, global semantic features and local lexical features are simultaneously extracted based on the clinical pretrained language model Bio_ClinicalBERT, and the two are fused through a heterogeneous textual feature fusion module, thereby enhancing the model's representation capability for the overall semantics of the chief complaints and local symptom-related keywords. In the speech branch, an Audio Spectrogram Transformer (AST) is used to extract temporal acoustic representations from speech, thereby supplementing expressive speech information that is difficult to capture from textual records. In the preliminary classification stage, the text and speech branches output initial classification results through their respective classification modules, allowing the two modalities to perform symptom evaluation independently. In the final classification stage, an adaptive gated decision fusion strategy is designed to dynamically generate fusion weights according to the features of different samples. The initial classification results from the text and speech branches are then weighted and fused to obtain the final symptom classification results. Unlike simple feature concatenation or fixed-weight fusion, this strategy can adaptively adjust the contribution of the two modalities to the final decision based on sample differences, thereby enhancing the influence of discriminative information on the classification results and improving the classification stability of the model in complex chief complaint scenarios. Experimental results on a public medical dataset show that the DWDF-MSC achieves 82.43%, 87.44%, and 81.52% in accuracy, precision, and F1-score, respectively, outperforming most mainstream baseline models across all metrics. A comparison of multimodal fusion schemes further demonstrates that the proposed dynamic weight decision fusion achieves better classification performance than feature-level fusion. In the ablation study, the complete DWDF-MSC model achieves relative improvements of 4.25% and 7.60% in accuracy and F1-score, respectively, compared to the variant that only employed heterogeneous text feature fusion, thereby demonstrating the effectiveness of the speech branch and the adaptive gated decision fusion mechanism. The McNemar test results show that the p-values between the DWDF-MSC and the multiple comparison methods are less than 0.000 1, indicating that the differences in classification results between the DWDF-MSC and these comparison methods are statistically significant. The antinoise performance experiments demonstrate that the DWDF-MSC can maintain a relatively stable classification performance under different Signal-to-Noise Ratio (SNR) conditions. In summary, DWDF-MSC can effectively fuse textual and speech information from chief patient complaints, improve model classification performance, and provide a feasible multimodal method for intelligent symptom classification based on chief patient complaints.
-
Computer Engineering. 2026, 52(10): 317-327. https://doi.org/10.19678/j.issn.1000-3428.0260120
Multimodal Emotion Recognition in Conversations (ERC) integrates language, acoustic, and visual information to automatically identify emotions in dialogues, thereby enhancing naturalness and emotional understanding in human-computer interactions. However, existing methods have limitations in modeling the multilayer contextual dependencies of emotions. Multimodal feature fusion often introduces redundant information and noise, making it difficult to capture the uncertainty of emotions, thereby limiting the recognition of complex emotions. To address these issues, this study proposes a multimodal ERC model that combines hybrid encoding and fuzzy modeling. The model uses a hybrid encoding module to simultaneously capture the global dialogue context and local dependencies of emotions simultaneously, thereby enhancing the representational capability of emotional temporal features. Accordingly, a hierarchical gated fusion mechanism is introduced to perform dynamic weighted fusion on features from different levels and modalities, effectively suppressing redundant information and noise. In the emotion classification stage, a fuzzy neural network with linear equally spaced initialization models the boundaries of emotion categories using fuzzy membership functions to capture the uncertainty and fuzziness in emotional expressions. The experimental results show that the proposed model outperforms the baseline models for all metrics across the IEMOCAP, MELD, and CMU-MOSEI datasets. It achieves an accuracy of 72.67% for IEMOCAP, 67.37% for MELD, 54.96% for seven-class classification accuracy, and 86.78% for two-class classification accuracy on CMU-MOSEI, validating the effectiveness of the proposed model in multimodal emotion recognition.
-
Computer Engineering. 2026, 52(10): 328-339. https://doi.org/10.19678/j.issn.1000-3428.0252253
Large Language Model (LLM)'s factuality hallucination refers to the generation of content that conflicts with established real-world facts, significantly reducing model credibility and applicability in high-risk domains such as healthcare, law, and scientific research. Current methods for hallucination mitigation primarily depend on input optimization, supervised learning, or integration with external knowledge bases. However, these approaches exhibit limited generalizability, substantial dependence on extensive labeled datasets, and constraints in real-time scenarios, making it challenging to fundamentally improve the factual accuracy of LLMs. To address these limitations, this paper proposes a reinforcement learning-based framework incorporating semantic entropy as feedback to mitigate factuality hallucinations. Semantic entropy serves as a precise measure of uncertainty at the semantic level, enabling an accurate assessment of the model's confidence in its generated responses. By embedding semantic entropy into the reinforcement learning process as a reward signal, the model is encouraged to proactively avoid responses with a high likelihood of hallucination. Compared to traditional predictive entropy-based methods, semantic entropy more effectively distinguishes semantically equivalent expressions and enhances factual judgment capabilities without reliance on external knowledge sources. Experimental results show that the proposed method, while maintaining the richness and coherence of the generated content, can improve factual judgment accuracy by up to 5.7 percentage points and factual generation accuracy by up to 7.8 percentage points, compared to the best baseline model, significantly validating its superiority in factuality hallucination.
-
Computer Engineering. 2026, 52(10): 340-349. https://doi.org/10.19678/j.issn.1000-3428.0252412
In the home service scenario, multi-robot systems need to process natural language instructions issued by non-professional users, which poses higher requirements for automated task scheduling. Existing multi-robot scheduling methods present limitations in terms of task understanding, dependency management, and resource optimization. To address these issues, this paper proposes a Directed Acyclic Graph (DAG)-Large Language Model (LLM) scheduling method to achieve full process automation from natural language input to multi-machine collaboration. First, this method first utilizes the LLM combined with environmental information for semantic parsing and task decomposition and generates a set of subtasks with execution constraints through the Chain-of-Abstraction (CoA) mechanism. Second, based on the LLM, a DAG is automatically constructed between subtasks to replace traditional manual modeling processes and accurately represent task dependencies. Finally, a backtracking algorithm is used to match robot skills with subtask requirements, combined with an asynchronous execution strategy to improve execution, and dynamic scheduling is used to reduce waiting time while ensuring dependency order. To verify the effectiveness of this method, three types of household tasks with different complexities (including four sets of scenarios) are designed in an AI2-THOR simulation environment for comparative experiments. Experimental data show that, compared with SMART-LLM and AutoTAMP, the DAG-LLM improves the task success rate by 41.4 and 57.7 percentage points, respectively, and reduces the running time by 24.3% and 6.3%, respectively. The results of an ablation experiment further demonstrate that task dependency modeling and asynchronous execution mechanisms play a crucial role in improving the system performance. The proposed method does not require manual involvement in task decomposition and dependency modeling, and it is suitable for the efficient collaborative scheduling of multi-robot intelligent agents in natural-language-driven application scenarios such as homes.
-
Computer Engineering. 2026, 52(10): 350-362. https://doi.org/10.19678/j.issn.1000-3428.0252354
With the exponential growth of information dissemination on social media platforms, false news detection has become a critical challenge in information authenticity verification. Existing methods focus on single-modal semantic analysis, which not only inadequately models cross-modal semantic contradictions in multimodal news but also suffers from limited interpretability due to the absence of explainable auxiliary information. To address these issues, this study proposes a multimodal fake news detection framework that utilizes a vision-language large model. The framework introduces the following innovations: 1) utilization of the vision-language large model Qwen2.5-VL to reason for news content and generate multimodal description sets that enhance interpretability; 2) design of a multi-granularity co-attention mechanism to achieve cross-modal feature alignment across textual semantics, visual features, and auxiliary descriptions. News-prompting templates are designed to guide Qwen2.5-VL in extracting key objects, scene elements, and contextual semantic enhancements from the news, thus generating explainable auxiliary decision-making evidence. Based on the co-attention layers, the multi-grained co-attention fusion mechanism employs hierarchical feature interactions to capture latent fake patterns in multimodal news within high-dimensional semantic spaces. In addition, experiments are conducted on three multimodal fake news datasets, namely, Weibo, GossipCop, and Pheme, and the results show that the accuracy rates reach 90.4%, 88.1%, and 86.6%, respectively.
-
Computer Engineering. 2026, 52(10): 363-374. https://doi.org/10.19678/j.issn.1000-3428.0070706
Cyber Threat Hunting (CTH) enables rapid response to attack events through proactive discovery of attack clues and malicious evidence. Although existing CTH algorithms can search extensive information sources, they face challenges in real-world scenarios owing to insufficient prior knowledge and sparse feedback. To address these problems, this paper proposes a CTH algorithm, RE-HUNTER, based on Large Language Model (LLM) and Reinforcement Learning (RL). To address the lack of prior knowledge, this algorithm constructs a contextual vector database and leverages domain expertise from LLM and unstructured knowledge from cyber threat intelligence for cold-start decision-making to initialize RL weights. To address sparse feedback, this algorithm improves the Monte Carlo Tree Search (MCTS) algorithm by introducing a recursive update mechanism and method similarity to enhance the feedback on the execution results of both entities and methods. Experiments conducted on 186 real-world attack cases demonstrate that this model significantly improves the search efficiency compared with the current state-of-the-art baseline methods. Within the 0—2 000 step range, the average recall rate achieves a relative improvement of 18.24%. Notably, in the 0—250 step cold-start scenario, the average recall rate achieves a relative improvement of 86.28% compared to that achieved by the best baseline method. Furthermore, the ablation experiments indicate that each component of the proposed algorithm positively contributes to the overall performance, effectively reducing the cost of CTH.
-
Computer Engineering. 2026, 52(10): 375-389. https://doi.org/10.19678/j.issn.1000-3428.0070735
In recent years, large models have become one of the most important tools for humans. Prompts play a crucial role as bridges of communication between humans and large models. Although high-quality prompts can fully unleash the potential of large models, their design requires both specialized skills and substantial resources, leading to the emergence of prompt trading markets and licensing service models. However, treating prompts as commodities presents three main challenges: 1) Loss of value of prompts owing to them being easily copied and disseminated once the prompt text is leaked; 2) Lack of objective standards and pricing systems to ensure prompt quality; 3) Difficulty experienced by users in holding providers of false prompts accountable, and rights of providers being affected by the actions of centralized platforms. To address these challenges, this study proposes PromptDEX, a blockchain-based service platform for large model prompts. Through a prompt rental mechanism based on smart contracts and service templates built on LangChain, the platform enables direct interaction between providers and demanders, thereby reducing the impact of centralized institutions on interests and privacy of the participants. All transaction records are transparently stored in the blockchain to ensure security, reliability, and accountability. Additionally, a dynamic pricing mechanism based on demand-side ratings is implemented. The experimental results show that the additional cost incurred by integrating the blockchain is negligible, as calculated in terms of fractions of a cent. Providers can build API services with only approximately 10 lines of code, requiring minimal network bandwidth and demonstrating strong feasibility and practicality.
-
Computer Engineering. 2026, 52(10): 390-406. https://doi.org/10.19678/j.issn.1000-3428.0260022
With the rapid development of the internet and social media, the speed of information generation and dissemination has reached an unprecedented level. The proliferation of misinformation, rumors, and other misleading content has become increasingly prominent, posing significant threats to the order, harmony, and stability of social governance. In rumor detection, a low proportion of rumor samples leads to data imbalance, and existing text augmentation techniques struggle to enhance detection performance owing to their lack of specificity to rumor styles and low generation quality. Additionally, although pretrained language models capture global dependencies in text, they often fail to focus on the key local features of rumors. To address these challenges, this study proposes a rumor detection framework based on large-model data augmentation and multi-granularity feature fusion. First, a rumor generation method integrating a rumor-style lexicon and Large Language Model (LLM) is proposed. Based on publicly available rumor datasets, a style lexicon is constructed to guide LLM in generating semantically coherent and rumor-style consistent minority class samples. This approach alleviates data imbalance while ensuring the quality of the augmented samples. Second, this study introduces a Multi-Granularity Contextual Feature Extractor (MGCFE), which combines the strengths of pretrained language models with disentangled attention mechanisms to capture global dependencies and focus convolutional sublayers on local features. This enables the simultaneous capture of long-distance logical associations and fine-grained linguistic clues in rumor semantics, effectively mitigating the inherent limitations of pretrained models in capturing key local features. The experimental results demonstrate that the proposed detection method achieves accuracy rates of 82.24% and 93.91% for BuzzFeed and PolitiFact datasets, respectively.
-
Computer Engineering. 2026, 52(10): 407-417. https://doi.org/10.19678/j.issn.1000-3428.0252069
Vision-and-Language Navigation (VLN) tasks aim to guide an agent to move to a target location in a 3D or real-world environment based on language instructions. However, the traditional end-to-end VLN algorithms have certain limitations. When an erroneous action occurs during navigation planning, the agent tends to enter incorrect paths, resulting in an inability to continue following the instructions or exploring unnecessary areas. To address this issue, an agent called Nav-Explore is proposed, which is based on a Large Language Model (LLM) and an exploration module. The proposed agent leverages the reasoning capabilities of the LLM to predict the next action according to the language instructions and current visual information and uses an exploration module to balance exploration and utilization. The exploration module employs an ε-greedy strategy for toggling between the normal navigation and exploration modes. When the random probability is below ε, the agent explores possible future paths to assess the feasibility of the next actions, thus avoiding incorrect decisions. If the probability exceeds ε, it directly uses the output of the LLM for navigation. This modular design enables Nav-Explore to effectively enhance navigation success rates and improve the generalization ability of the agent in unseen environments. The experimental results demonstrate that Nav-Explore achieves superior performance on two outdoor VLN benchmark datasets, Touchdown and Map2seq, significantly increasing navigation success rates. Furthermore, Nav-Explore exhibits strong generalization capabilities, effectively completing navigation tasks in different environments.
-
Computer Engineering. 2026, 52(10): 418-428. https://doi.org/10.19678/j.issn.1000-3428.0070444
With the widespread adoption of microservices in digital service networks, the large number of service nodes and the complexity of call graphs in these networks present significant challenges for operational management. Although distributed tracing technology has made significant progress, it still has limitations such as the need to invade source code and reliance on specific middleware. These limitations lead to insufficiencies in the accuracy and completeness of tracing chains and affect the reliability of downstream analysis tasks based on observability data. To address these issues, this paper proposes a Trace-Metric oriented HEterogeneous Dynamic Graph neural network (TM-HEDGE). First, a heterogeneous dynamic directed acyclic graph is constructed by incorporating metric data. Subsequently, an intra-snapshot heterogeneous attention encoder and an inter-snapshot Transformer encoder are used to learn node-heterogeneous spatiotemporal representations. Finally, missing tracing chains are completed through link completion, thereby realizing tracing reconstruction. Experiments are conducted to evaluate the tracing reconstruction performance. The results show that TM-HEDGE improves the accuracy by 5.22% on average compared with existing state-of-the-art link completion model on three public datasets, which significantly enhances the completeness and accuracy of tracing chains in digital service networks.
-
Computer Engineering. 2026, 52(10): 429-441. https://doi.org/10.19678/j.issn.1000-3428.0252087
In the fields of network communication and traffic management, rapid and accurate identification of heavy flows is of great significance for tasks such as traffic congestion control and malicious traffic monitoring. However, extremely high data transmission rates in real-world network environments render heavy flow detection exceptionally complex and challenging. Most existing heavy flow detection methods rely primarily on single-dimensional statistical data; typically, these methods obtain traffic statistics and perform analyses based solely on estimated flow packet counts. The limitation of such an approach lies in its neglect of critical information from other potential dimensions, such as the distribution characteristics of packet inter-arrival times, which may be crucial in accurately identifying heavy flows. To address these issues, this paper proposes a heavy flow detection algorithm called IntervalSketch. By incorporating two-dimensional traffic features—the estimated flow size and packet interarrival time distribution—the algorithm optimizes the protection of heavy flows and the replacement strategy for mouse flows. Specifically, IntervalSketch effectively distinguishes heavy flows from mouse flows by leveraging packet inter-arrival time characteristics, thereby significantly improving the detection performance under low-memory conditions. To validate the accuracy and effectiveness of IntervalSketch, two real-world network traffic datasets, CAIDA and MAWI, are used. The experimental results demonstrate that IntervalSketch exhibits significant advantages across various memory configurations and traffic distributions. Compared with existing methods, it maintains a higher detection accuracy under memory-constrained conditions, achieving an F1 value up to 2.4 times that of state-of-the-art approaches.
-
Computer Engineering. 2026, 52(10): 442-454. https://doi.org/10.19678/j.issn.1000-3428.0252127
In malicious encrypted traffic classification, models expand the richness of discriminative representations by increasing traffic feature dimensions. However, issues such as the mismatch between the selected models and the data characteristics of malicious encrypted traffic, as well as insufficient feature selection still persist. Moreover, research and discussion on the data characteristics of encrypted traffic is lacking. To address these challenges, this study proposes a multi-representation fusion-based classification model for Internet of Things (IoT) malicious encrypted traffic classification. An abstract representation learning model is employed to learn packet-level byte-association representations and statistical representations of traffic sessions. Additionally, a plaintext representation learning model is used to learn session connection representations from unencrypted plaintext. Finally, the two models are fused based on the confidence scores of the abstract representation learning model to obtain the final malicious traffic classification results. The superiority of the proposed model is validated by comparing it with seven baseline models. The results show that the proposed model achieves an F1 value of 0.769 4, representing a significant improvement over all baseline models. Furthermore, to verify the compatibility of each module with traffic representation learning and the complementarity among discriminative representations contained in the selected features, ten variant models based on different inputs and model architectures were generated for comparison. The results demonstrate that the proposed model achieves superior detection performance, thereby proving the compatibility of the model architecture and complementarity among the representations.
-
Computer Engineering. 2026, 52(10): 455-464. https://doi.org/10.19678/j.issn.1000-3428.0252187
The handling of computation-intensive tasks is challenging for User Devices (UD) with limited computing resources. Edge computing addresses this problem by extending computing resources to the network edge, and one of its key functions is rational computational offloading. However, coordinating the computing resources of numerous edge nodes for task offloading while ensuring data security throughout the process is a major challenge. To address this issue, this paper proposes a Deep Reinforcement Learning (DRL)-based secure task offloading method. First, an edge computing network model is constructed, and a variable security protection mechanism is designed to adaptively safeguard data. The edge computing network model and task objectives are formalized and transformed into a Markov Decision Process (MDP). Finally, a DRL method based on a penalized action space is proposed to derive the optimal task offloading strategy. Simulation results demonstrate that the proposed scheme can reduce latency and energy consumption costs while maintaining security protection, and it consistently achieves a zero task loss rate.
Volume 52 No 10
Superintended by:
China Electronics Technology Group Corporation
Sponsored by：
East-China Institute of Computer Technology
Shanghai Computer Society
Edited & Published by：
Editorial Office of Computer Engineering
Chief Editor：Wu Jiangxing
Executive Chief Editor:
Hu Bing
Tel：021-67092514
E-mail: ecice06@ecict.com.cn
CN: 31-1289/TP
ISSN: 1000-3428