In the modern industrial sector, the perception and analysis of text data are essential for promoting intelligent manufacturing and optimizing production processes. However, industrial text data are typically characterized by high specialization, diversity, complexity, and annotation costs, making traditional large-scale annotation methods unsuitable. Existing few-shot Named Entity Recognition (NER) methods often use prototypical networks to classify entities, where the prototype is the average of the features of all samples belonging to the same category. However, these methods are highly sensitive to the support set data and prone to sample-selection bias. To address this, a few-shot NER model based on Distribution Calibration called DC-NER is proposed. A two-stage approach for span detection and entity classification is adopted for the few-shot NER task. In the first stage, an entity extractor is used to select candidate entities. In the second stage, a type discriminator assigns these entities to predefined categories. To address the challenge of capturing the true class distribution with a small number of labeled samples in few-shot NER, which leads to inaccurate class prototype calculations, a method is proposed that leverages useful information from the source domain data to calibrate the class distribution in the target domain. Thus, more samples are generated based on the refined class distribution to construct more accurate prototypes, thereby improving the performance in few-shot NER tasks. Experiments on both the in-domain dataset Few-NERD and the cross-domain dataset Cross-NER reveal that DC-NER significantly outperforms the comparative models in terms of F1 value, validating its effectiveness for few-shot NER.
Bayesian Optimization (BO) is a powerful tool for sequential optimal design in formulation development. However, conventional BO is sensitive to initial samples, easily falls into a local optimum, and cannot dynamically adapt to changes in the process easily. By combining the global search capability of Reinforcement Learning (RL) with the efficient exploration strategy of BO, this study proposes a sequential formulation optimization design method based on a Deep Deterministic Policy Gradient (DDPG) to achieve adaptive formulation optimization. By leveraging real-time feedback from the data, this approach dynamically adjusts the optimization strategy, enabling efficient global optimization starting from random initial points. It effectively addresses challenges such as limited data, inaccuracies in surrogate models, and measurement noise. During the optimization process, the method employs a Gaussian Process (GP) to construct a surrogate model of the objective function and adopts an inner-outer dual-loop framework to enhance optimization efficiency. In the inner loop, the actor network generates candidate actions, and the Upper Confidence Bound (UCB) strategy is used to select optimal experimental points. In the outer loop, the experimental strategy is dynamically adjusted based on historical experience, and the surrogate model is updated using new observations. The results from simulation cases and wine blending experiments demonstrate that the proposed method outperforms traditional optimization methods in terms of optimization efficiency, convergence speed, and quality of the final solution. Moreover, it exhibits a strong generalization ability in new environments, providing a novel solution for complex formulation optimization problems.
With the frequent occurrence of terrorist attacks, crowd evacuation path planning in indoor public places has received increasing attention. This paper proposes a path planning method based on the Proximal Policy Optimization (PPO) algorithm to improve the efficiency and safety of pedestrian evacuation. The indoor terrorist attack scenario is described, and static obstacles, idle locations, dynamic obstacles, exits, and pedestrians in indoor public places are modeled using a cellular automata model. Subsequently, a feature construction method based on distance information is proposed to construct pedestrian features that represent the escape difficulty of evacuation paths, including shortest path features and safe path features, by combining the distance from pedestrians to exits in non-threatening environments and threat-facing scenarios. Finally, by describing the evacuation path planning problem as a reinforcement learning problem, a reward function based on evacuation efficiency, death penalty and successful escape reward is designed. Based on the feedback from the real-time environment, an evacuation strategy is provided to pedestrians, which in turn realizes the overall optimization of the escape path by the PPO algorithm. Compared with existing field methods, the proposed method can improve the efficiency and safety of crowd evacuation in different simulation scenarios, particularly in complex and high-density environments. The effectiveness of the shortest and safe path features is verified through ablation experiments.
Electric power load forecasting is a crucial step in optimal dispatching of power grids. However, when confronted with complex dynamic environments and incompletely stationary load characteristics, existing purely data-driven methods and strategies incorporating domain knowledge insufficiently capture long-term trends and periodic patterns, which affects forecasting accuracy and robustness. To address this issue, an electric power load forecasting model based on the integration of domain prior knowledge, named DPK-ELF, is proposed. This model utilizes a prior knowledge extraction module to thoroughly analyze the dynamic behavioral characteristics of time-series data, constructs domain prior knowledge specific to the given data, and employs a dynamic piecewise stacking average smoothing method to extract prior trends in electric power loads. The prior-trend decomposition module decomposes the electric power load series into prior smoothed trends and residual local random fluctuations, which are then combined with the PatchTST data-driven model for forecasting. Soft constraint optimization techniques are adopted during the model training phase using domain prior knowledge as boundary constraints in the loss function to enhance the robustness of the model. Experiments on four publicly available electric power load datasets reveal that DPK-ELF outperforms comparison models such as PatchTST, DLinear, Autoformer, and Informer across three key performance metrics: Mean Squared Error (MSE), Mean Absolute Error (MAE), and Relative Squared Error (RSE). In the Australian electricity price and power load dataset, compared with the PatchTST model, DPK-ELF achieves improvements of 28.31%, 19.57%, and 14.94% in MSE, MAE, and RSE, respectively. In the PDB power demand dataset, compared with the PatchTST model, DPK-ELF achieves improvements of 12.25%, 7.77%, and 6.29% in MSE, MAE, and RSE, respectively. These results demonstrate the significant advantages of the DPK-ELF model in improving forecasting accuracy.
The security of classical public-key cryptosystems, which operate based on the computational intractability of integer factorization and discrete logarithm problems, is profoundly threatened by the rapid development of quantum computing. Post-Quantum Cryptography (PQC) has been proposed to improve security against quantum attacks. Lattice-based cryptography is widely regarded as one of the most promising PQC method because of its proven security, efficient implementation, and functional versatility, and it has achieved significant progress in both algorithmic design and engineering implementation in recent years. This paper presents a systematic survey of the status of research on lattice-based cryptographic algorithms. First, the fundamental theory of lattices and their core hard problems are discussed. Next, the technical characteristics and performance of encryption and key encapsulation mechanisms, digital signature schemes, and key exchange protocols constructed based on lattice hard problems are reviewed and compared. Subsequently, the principles and optimization paths of two key technologies for lattice cryptography implementation, namely polynomial multiplication using the fast Number Theoretic Transform (NTT) and pre-image sampling algorithms, are discussed. Finally, considering the PQC standardization process of the National Institute of Standards and Technology (NIST), an outlook on future research directions for lattice-based cryptography is presented. The research directions include the development of more "succinct" fundamental hard problems, strengthening security proofs, improving algorithm efficiency, and countering novel attacks. This review serves as a reference for promoting the practical deployment and further research of lattice-based cryptography.
In the digital era, complex interactions among modalities such as text, image, and audio have led to multimodal misinformation. The propagation speed and concealment levels of multimodal misinformation far exceed those of traditional unimodal misinformation, posing severe challenges to information security and social governance. However, research in this field is relatively scarce in China, and a comprehensive framework has yet to be established. Therefore, this paper provides a systematic overview of the research status and development trajectory of multimodal misinformation detection. Based on a clear understanding of the core concepts and task spectrum of multimodal misinformation detection, this review provides a detailed analysis of the characteristics of the datasets and evaluation metrics. It also analyzes the applicability and detection performance of different multimodal methods and models, such as SAFE, CAFE, CFFN, SSA-MFND, PSCC-Net, DGM4, CCN, SNIFFER, and KGAlign. Three core detection methods, namely cross-modal consistency, anomaly feature recognition, and external fact-driven approaches, are summarized. Furthermore, the interpretability and generalization robustness of multimodal misinformation detection are analyzed. With the rise of Large Vision-Language Models (LVLMs), their application to multimodal misinformation detection is continuously increasing. Various application scenarios, advantages, and limitations of LVLMs in this domain are discussed. Finally, future research directions in multimodal misinformation detection are outlined, aiming to provide insights and inspiration for further development in this field.
Previous studies on cross-view geolocalization have primarily focused on determining whether a query image accurately corresponds to a specific geographic location within a predefined gallery. However, this research paradigm often overlooks the extensive multiscale structural information inherent in the geographic space. To achieve more robust localization, a model must not only capture local architectural details but also understand the spatial relationships among targets reflected through building clusters and environmental features, thereby improving the localization accuracy across different spatial scales. To address these challenges, a multiscale cross-view geo-localization task is proposed and a Multi-Level Campus (ML-Campus) dataset is constructed specifically for this task. The ML-Campus dataset comprises multiview, multisource building images, each annotated with multiscale labels to reflect correlations and continuity across different spatial scales. Based on this dataset, an empirical evaluation of existing cross-view geo-localization methods is conducted, which is used as a benchmark to measure their performance in this context. To further enhance model performance, the proposed Cross-View HAPPIER (CV-HAPPIER) method is employed for training, which strengthens the model's feature representation capabilities across different spatial scales. Extensive experimental results on the ML-Campus dataset demonstrate that the CV-HAPPIER method significantly improves the spatial robustness of cross-view geo-localization retrieval ranking results.
Crowd escape behavior in public places is highly likely to cause serious public safety disasters. Traditional computer vision technology can detect a few characteristics of crowd escape behavior, but it is difficult to face complex dynamic visual scenes. To address this issue, based on the structural characteristics of the locust visual nerve, and leveraging the danger perception mechanism of the locust Lobula Giant Movement Detector (LGMD) as well as the mammalian retinal luminance adaptation mechanism, this paper proposes an Enhanced Crowd Escape Detection Neural Network (ECEDNN). First, the proposed neural network collects the luminance changes caused by crowd activities in the field of view. With the help of the mammalian retinal luminance adaptive mechanism, the visual response excitation is tuned to adapt to the lighting scene. Visual excitation and suppression are combined to filter background noise, and a center-surround mechanism is used to enhance motion edges. Finally, neural spike adaptive tuning is used to detect the burst escape behavior of the crowd and output strong membrane potential excitation. The experimental results show that ECEDNN can effectively detect and warn of crowd escape behavior in complex scenes, with an average accuracy of 98.90% on multiple video datasets. This work is involved the research of crowd activity detection inspired by biological visual perception mechanism, which can provide new ideas and methods for crowd behavior activity perception and anomaly detection in artificial intelligence.
Murals, an important part of cultural heritage, have received widespread attention in recent years for their digital protection and restoration. However, Super-Resolution (SR) reconstruction of mural images often faces challenges such as texture blurring and loss of original information. To address these issues, this study proposes a Reference-based Two-stage Mural Image Super-Resolution Reconstruction (RTMISR) method. First, a Multi-Scale Residual Feature Extraction Module (MRFEM) is employed to accurately capture the feature relationships between High-Resolution (HR) and Low-Resolution (LR) mural images, ensuring complete retention of LR image information and achieving an initial reconstruction of mural contours and partial details. Subsequently, a Mural Texture Feature Enhancement Module (MTFEM) is introduced, which utilizes a coarse-to-fine feature matching method to extract high-quality texture information from reference images and effectively integrate it into the reconstructed images to enhance texture detail representation. Moreover, to ensure the relevance and quality of the reference images, a Reference Image Filtering Module (RIFM) is designed to select reference images that are highly correlated with the target LR images. Experimental results on mural datasets show that, compared to representative SR methods such as SRGAN, MADNet, and ESRT, RTMISR achieves superior performance in objective metrics: for 2× SR, Peak Signal-to-Noise Ratio (PSNR) improves by an average of 2.83 dB and Structural Similarity Index Measure (SSIM) by 0.04; for 4× SR, PSNR improves by an average of 2.00 dB and SSIM by 0.02. In terms of subjective visual quality, RTMISR effectively retains the original information while enhancing the texture details of mural images, achieving a better balance between model complexity and reconstruction performance.
This study proposes TGMM, a large-scale code clone detection method based on parse trees and Graphics Processing Unit (GPU) acceleration, to address the limitations of existing tools in multilingual adaptation and large-scale analysis. The method employs a three-stage architecture for clone detection. First, it generates standardized parse trees based on lexical and syntactic rules of each programming language and then extracts subtrees that meet specific granularity requirements. Second, it simplifies these subtrees by pruning and eliminates nonfunctional differences through semantic equivalence transformations. Finally, it leverages GPUs to construct a global suffix array in parallel, enabling rapid similarity calculation of large-scale code blocks. The proposed method is evaluated in terms of clone detection efficiency and language scalability. On the public benchmark dataset BigCloneBench, TGMM achieves a precision of 97%, significantly outperforming seven mainstream tools compared, while reducing execution time by over 50% compared to the second-best tool and maintaining a recall rate similar to the compared tools across various clone types. In the language scalability test, TGMM successfully parses 19 out of 22 mainstream programming languages. Furthermore, by applying TGMM to conduct a multi-granularity clone analysis of the top 45 GitHub projects (covering nine programming languages), the study reveals significant differences in clone density across different languages, provides an in-depth analysis of the underlying causes, and thus offers practical and effective references for software maintenance.
With the development of multimedia technology, the difficulty of unauthorized forgery and dissemination of false information has significantly decreased. This can lead to several negative consequences. Effective content authentication algorithms are urgently required to ensure the authenticity and security of image content. Recently, perceptual image hashing has demonstrated excellent performance in the field of image authentication. However, existing algorithms are not ideal for processing images with a large proportion of text, and they cannot effectively cope with new content-preserving manipulations such as scribbling. Therefore, a Text—Picture Mixed Image (TPMI) content authentication algorithm based on perceptual hashing is proposed. The proposed algorithm adopts the image segmentation algorithm of ring partitioning and calculates the frequency and distribution characteristics of the Scale-Invariant Feature Transform (SIFT) keypoints within each ring. These features exhibit rotational invariance and effectively improve the anti-collision performance of the proposed algorithm. By obtaining keypoint information, the proposed algorithm performs well in terms of robustness against content-preservation manipulations, including irregular scribbles. A TPMI dataset is constructed to validate the performance of the proposed algorithm. Compared to some representative algorithms, this algorithm performs better in terms of perceptual robustness, anti-collision, and security. Partial tampering with images can effectively identify tampered images that are similar to the original image. Additionally, experiments on scribble attacks are conducted, and the results show that the proposed algorithm can effectively identify such attack images.
Low-Light Image Enhancement (LLIE) is crucial in computer vision by restoring rich visual information from corrupted low-light images. However, existing LLIE methods often suffer from color bias due to color space and typically fail to balance denoising and color fidelity within a single-stage framework. To address these challenges, this study introduces a novel Dual-stage HVI-based Transformer Network (DHTNet) for LLIE. DHTNet significantly improves the quality of low-light images by decoupling I and HV maps, enabling their independent yet synergistic optimization within the HVI color space. In the first stage, a hierarchical Transformer network equipped with an Adaptive Guidance Interaction Module (AGIM) models long-range dependencies between I and HV features to achieve global noise suppression and accurate color calibration. In the second stage, the Multi-Scale Enhanced Synergistic Attention (MESA) module enhances localized color and detail representation through synergistic optimization across the I and HV branches. This dual-stage framework effectively addresses the limitations of existing LLIE methods by retaining complex image details while enhancing visual realism. Experimental results show that DHTNet achieves the highest Peak Signal-to-Noise Ratio (PSNR) on both the SICE and SID datasets, outperforming the second-best model by 0.717 dB and 1.897 dB, respectively. Furthermore, DHTNet attains PSNR values of 28.756 dB, 24.683 dB, and 25.950 dB on the LOLv1, LOLv2-Real, and LOLv2-Synthetic datasets, respectively, consistently outperforming existing models such as Retinexformer and CIDNet.
Bi-triangle enumeration in bipartite graphs is a core operation for graph analysis tasks such as local Clustering Coefficient (CC) computation. As the scale of real-world bipartite graph data continues to increase, the data volume has exceeded the processing capacity of a single machine, necessitating efficient enumeration through distributed algorithms. However, existing distributed graph partitioning algorithm, called GP, suffer from issues such as large subgraph combination data volume, message overload, and redundant enumeration. To address these problems, two optimization algorithms based on the topological characteristics of bi-triangles are proposed and the graph partitioning strategies are customized accordingly. Method 1 treats a bi-triangle as composed of three wedge structures and generates subgraphs using wedge groups as the basic unit. It introduces a subgraph combination construction mechanism based on the concatenation of A-type and V-type wedge groups, significantly reducing the number and data size of subgraph combinations, and ultimately enumerates bi-triangles via wedge triplets. Additionally, to address message overload and redundant enumeration, Method 1 proposes a subgraph-reading mechanism based on a distributed storage system and a deduplication mechanism based on vertex ordering. Method 2 considers a bi-triangle as composed of two zedge structures. It first performs initial graph partitioning using wedge groups as the basic unit, then completes the second partitioning through a construction and restoration mechanism of "compressed zedge groups", ultimately enumerating bi-triangles via zedge pairs, achieving lower computational complexity than Method 1. Experimental results demonstrate that, compared with the GP algorithm, Method 1 processes only 0.5% of the subgraph combination data on average and the enumeration time is less than 2.2% of the original, whereas Method 2 processes approximately 3.3% of the data on average and the time is compressed to within 1.0% of the original.
In distributed computing frameworks, inefficient data transfer in the Shuffle phase has become a key bottleneck in data connectivity. Existing methods have certain limitations in dealing with table joins, such as broadcast joins and Hash joins in Spark are both susceptible to data skewing, which makes the load between nodes unbalanced. Aiming at this problem, the paper focuses on joining aggregated queries, and proposes a table joining method based on lattice structure: by precomputing the storage table partition data in the form of lattice structure, and utilizing the convex set property of equivalence class, i.e., the data cells containing the upper bound of equivalence class and contained by the lower bound of equivalence class, whose aggregation values are equal to the aggregation values of equivalence class, so as to realize the quick matching and calculation. Since the query data cells as a compressed form of basic table data, the data size and skew are more concise and uniform, the article uses the query data cells instead of table data to perform data transfer and connection, which greatly reduces the data Shuffle and computational complexity. The method proposed in the paper has been implemented in Spark, and experiments based on the TPC-H dataset show that: the method of the paper reduces the data Shuffle by about 45.06% in large dataset scenarios, meanwhile, the workload among the nodes is more balanced compared to the benchmark method, and the query response time is shortened by 14.23% on average.
This paper studies operator-level reconstruction for the structural mismatch between YOLOv11 and Ascend Neural Processing Unit (NPU). The Spatial Pyramid Pooling-Fast (SPPF), C3K2, and C2PSA modules are optimized without changing network semantics or model scale. Three Ascend C operators are designed: the SPPF operator uses on-chip data loop and halo cache to reduce redundant global-memory traffic in multi-stage pooling; the C3K2 operator integrates multi-core task assignment and multi-queue asynchronous pipelining to reduce fine-grained kernel launch overhead; and the C2PSA operator reconstructs attention communication through a parallel reduction-broadcast primitive. On an Ascend 910B NPU, the complete reconstruction reduces the training time per epoch by 23.2% and improves the training throughput by 27.6% on the COCO dataset. The results show that matching Ascend on-chip memory, asynchronous queues, and multi-core synchronization mechanisms improves the training execution efficiency of key YOLOv11 modules and keeps inference performance stable. It can provide verifiable operator mapping schemes for the deployment of complex object detection networks on the Ascend platform.
Existing time series forecasting methods fail to fully consider the spatiotemporal dependencies among variables, which hinders the improvement of forecasting accuracy. Spatial modeling methods based on Graph Neural Networks (GNN) also have limitations in graph structure construction. Specifically, static predefined graphs struggle to capture the dynamic interaction characteristics among variables, and adaptive graph structure learning is influenced by parameter initialization, rendering the method prone to being trapped in local optima. To address these issues, a multivariate time series forecasting method that combines spatiotemporal information and the Kolmogorov—Arnold Networks (KAN) is proposed. In the spatial dimension, a graph structure learning module is designed. It utilizes the Pearson correlation coefficient to establish the initial adjacency matrix of variables, introduces learnable parameters to dynamically adjust and optimize the graph structure, and effectively captures the spatial dependencies among variables by stacking multiple layers of graph convolutions. In the temporal dimension, the method combines the multi-head self-attention mechanism and Gated Recurrent Unit (GRU) to extract temporal dependencies in different subspaces, while capturing both local temporal patterns and global key information. To further enhance the representational capacity of the model, the KAN is used instead of the traditional Multi-Layer Perceptron (MLP), enabling nonlinear fusion of spatiotemporal features through learnable activation functions. Experimental results on seven benchmark datasets demonstrate that the proposed model achieves an average reduction of 36.9 percentage points in Mean Squared Error (MSE) and 24.8 percentage points in Mean Absolute Error (MAE). The generalization performance of the model is validated via testing using the Australian electricity load dataset. Compared with other mainstream models, this model can accurately capture the dependencies among variables and effectively fuse spatiotemporal features, thereby improving forecasting accuracy and robustness.
The efficient utilization of resources across nodes in a heterogeneous cluster is pivotal for enhancing the overall resource utilization of the cluster. One of the primary functions of Kubernetes, the preferred system for container orchestration, is to schedule pods onto appropriate nodes. A well-designed scheduling strategy can reduce resource fragmentation, enhance the utilization of node resources, and minimize job waiting times. In the context of resource scheduling within heterogeneous clusters, the default Central Processing Unit (CPU)- and memory scheduling algorithms of Kubernetes have become inadequate for addressing the demands of multi-dimensional resource job scheduling. To address this challenge, a multi-criteria scheduling algorithm integrating an Analytic Hierarchy Process (AHP) and binpacking scheduling concepts is proposed. This algorithm considers not only traditional CPU and memory resource indicators but also incorporates various other resource indicators such as storage, GPU-like accelerator cards, and images, enabling a more comprehensive resource assessment. It also combines subjective user decisions with quantified objective node resource constraints to select the most suitable node from multiple candidates for job scheduling. A custom scheduler is designed based on the Kubernetes scheduling framework. The use of a custom scheduler with a multi-criteria scheduling algorithm can more efficiently utilize node resources and reduce the job waiting time caused by resource fragmentation. Experiments on mixed deployments of different types of jobs in heterogeneous clusters demonstrate that the multi-criteria scheduling algorithm can effectively increase cluster resource utilization and reduce Pod scheduling wait times by 49.79%. This provides an important reference for optimizing scheduling strategies for future large-scale cloud platforms and data centers.
The demand for continuous time-series data sharing in Internet of Things (IoT) scenarios has been increasing. Traditional Certificateless Signcryption (CLSC) schemes based on bilinear pairings or elliptic curves offer limited security under quantum computing threats and incur high computational and communication overheads when repeatedly signcrypting multiple data segments. To address these issues, this paper proposes a post-quantum lightweight certificateless hybrid signcryption scheme tailored for time-series data sharing. First, a bidirectional hash chain-driven symmetric key generation and management mechanism is designed to enable rapid key generation for multiple consecutive data segments while ensuring forward and backward security. Subsequently, a lightweight hybrid signcryption scheme with conditional privacy protection is constructed using a module lattice framework. Combined with an offline/online separation that shifts part of the computation ahead of time, the scheme achieves post-quantum security while significantly reducing the burden on end devices and communication links. The scheme is proven to satisfy IND-CCA confidentiality and EUF-CMA unforgeability in a random oracle model. Prototype implementation and evaluation on resource-constrained terminal devices show that, the proposed scheme achieves significant advantages: on average, the end-side overall latency and transmission volume can be reduced to approximately 40% of those of comparable post-quantum schemes. The advantages of this scheme become even more pronounced in long-duration continuous time-series data scenarios. It effectively balances post-quantum security, lightweight design, and scalability, making it well-suited for time-series data sharing applications in next-generation intelligent IoT.
Existing Knowledge Graph Completion (KGC) methods fail to effectively distinguish semantic differences among paths at different levels and cannot dynamically adjust relation representations by fully leveraging neighborhood context information. Consequently, contextual semantics remain insufficiently understood. To address these issues, this paper proposes a KGC model integrating relational context semantics, named RCSKGC, for enhancing the semantic representation of learned paths and neighborhood information. First, the local and global encoding of multi-hop paths at different levels is performed using a Bidirectional Gated Recurrent Unit (Bi-GRU) and an attention mechanism to extract effective path information. Meanwhile, the fine-grained semantic features of the path information at different levels are further enhanced through relational embedding contrastive learning. Second, a dual-attention mechanism and dynamic weighting strategy are employed to capture the neighborhood-level information of relations and learn their semantics through a "neighborhood—entity—relation" approach. Finally, the two types of relational representations are aggregated and used as inputs for the decoder to perform KGC. The results of experiments on the FB15k-237 dataset demonstrate that, compared to the best results of the baseline methods, RCSKGC achieves improvements of 1.4, 0.8, 1.3, and 2.1 percentage points in the Mean Reciprocal Rank (MRR), Hits@1, Hits@3, and Hits@10, respectively. Further, on the WN18RR dataset, RCSKGC shows improvements of 0.8 and 1.0 percentage points in MRR and Hits@3, respectively. These results validate the effectiveness of the proposed method.
The Hypertext Transfer Protocol (HTTP), the core infrastructure of Internet communication, relies on the collaboration of multiple servers in its modern communication model. If the servers in the processing chain do not strictly follow the protocol specifications or have differences in semantic interpretation, systemic semantic inconsistency problems will occur, leading to security threats such as access control strategy bypass, multi-host issues, request smuggling, and cache pollution. Differential fuzz testing analyzes semantic inconsistency problems by observing the differences in the processing results of messages are processed across different servers. To solve the problems of an inaccurate field set range, low mutation efficiency, and single observation dimension in existing tools, this paper proposes an improved differential fuzz testing method. First, a message construction method based on key headers is used to select core fields and streamline the test space. Second, a mutation method based on field semantics is adopted, combining semantic classification and vulnerability characteristics to design mutation operations, thereby enriching the test cases. Third, an extended message analysis method is employed to expand the scope of message analysis to both request and response messages, fully observing the communication process and covering existing scenarios of semantic inconsistency problems. Finally, tests are conducted on 7 commonly used servers, identifying 18 types of server processing differences and verifying 9 pairs of combinations that exhibit semantic inconsistency problems. Compared with similar tools such as t-reqs, the proposed method reduces the size of the test set by an order of magnitude, increases the average proportion of valid test cases by 12.67 percentage points, discovers two additional types of difference problems under the same observation conditions, and expands the test scope to cover all four scenarios of current semantic inconsistency problems.
