Table of Content

27 October 2025, Volume 0 Issue 10

Previous Issue   

Lightweight Model for Recognizing Abnormal Behavior in Factory Personnel Based on Deep Learning

LIU Longen1, SHI Dongxiang2, SHAN Baoming1, ZHANG Fangkun1, XU Qilei1

2025, 0(10):  1-6.  doi:10.3969/j.issn.1006-2475.2025.10.001

Asbtract ( 94 )   PDF (3864KB) ( 115 )  

References | Related Articles | Metrics


Abstract: This paper proposes an improved lightweight network for recognizing abnormal behavior among factory personnel based on YOLOv5， to addressing challenges such as complex backgrounds and limited computational resources. This network integrates Omni-dimensional Dynamic Convolution （ODConv） and the Explicit Visual Center Block （EVCBlock）， resulting in improved detection performance while reducing parameter computation. The ODConv module is introduced in the neck network to  enhance the model’s adaptability to complex factory environments and decrease the number of model parameters， while the EVCBlock module is added at the end of the backbone network to improve the detection accuracy of the model and compensate for accuracy loss of model caused by the reduction of parameters. The Normalized Wasserstein Distance （NWD） loss is constructed to optimize the model training process and enhance the model’s detection performance on small targets. Several enhanced detection models are constructed based on existing lightweight methods to compare detection accuracy and parameter count. Results demonstrate that the proposed lightweight recognition model has fewer parameters while maintaining high detection accuracy compared with the existing methods. Compared with the original model， the mAP of the detection model built in this paper increases by 3.2 percentage points and GFLOPs decreases by 2.2. This work is of guiding significance to realize rapid detection and accurate identification of factory personnel’s abnormal behavior in industrial production scenarios.



Fusion of Spatial Information for YOLOv7 Traffic Sign Detection

SHI Hongyu, ZHANG Zheyu, DU Wen, LI Yi

2025, 0(10):  7-13.  doi:10.3969/j.issn.1006-2475.2025.10.002

Asbtract ( 52 )   PDF (2423KB) ( 73 )  

References | Related Articles | Metrics

Abstract: During the detection process of traffic signs， due to the influence of weather and light intensity， problems such as false detections and missed detections occur during detection. To solve this problem， a traffic sign detection algorithm combining spatial information is proposed. Firstly， coordinate convolution is used on network to enhance sensitivity of the network to coordinate position information. Additionally， the incorporation of a coordinate attention mechanism into the backbone features enables better focus on spatial location information at fusion points. Moreover， the feature fusion process utilizes a multi-scale weighted network and pyramid pooling， leveraging weighted calculations and skip connections to enhance semantic information fusion between low-level and high-level layers. Lastly， the adoption of the SIoU loss function enhances target positioning accuracy. The experimental results on the CCTSDB2021 and GTSDB datasets demonstrate that this method achieved mean Average Precision (mAP) values of 84.9% and 98.5% respectively. Compared with mainstream detection models， it shows significant improvement—exceeding the original model by 5.39 percentage points and 1.67 percentage points—thus enhancing the detection accuracy of traffic signs.



Sublingual Vein Image Segmentation Based on HRNetV2 Model with Variable Kernel Convolution

JIANG Dongmei1, YANG Nuo2 , CHEN Renming2, DONG Changwu3, PENG Chengdong2, 4

2025, 0(10):  14-19.  doi:10.3969/j.issn.1006-2475.2025.10.003

Asbtract ( 62 )   PDF (2984KB) ( 48 )  

References | Related Articles | Metrics


Abstract: The existing analysis of sublingual vein often uses convolutional neural network （CNN） image classification methods or image segmentation methods to extract. But there is a problem of low accuracy in extracting details of meridians. Therefore， an improved HRNetV2 high-resolution semantic segmentation network algorithm is proposed to extract sublingual meridians. Adopting a high-resolution HRNetV2 network structure， the outputs of sub network structures from high to low resolution are connected in parallel to form multi-scale fused feature maps with higher spatial accuracy， improving the problem of loss of detailed information in sublingual veins. In addition， AKConv， convolutional kernel with arbitrary sampled shapes and arbitrary number of parameters instead of ordinary convolution can improve the convolution’s adaptability to change pulse structure and reduce the problem of under segmentation. The algorithm is validated through data extraction on the tongue image open platform of Anhui university of Traditional Chinese Medicine （TCM） cloud diagnosis technology， with pixel accuracy（PA）， mean pixel accuracy（mPA）， and mean intersection over union（mIoU） of 95.28%， 92.33%， and 93.42%， respectively， which is superior to the Mask-RCNN model， U-Net models， and HRNetV2 model. The improved HRNetV2 method has high accuracy in segmenting sublingual vein images， providing a new method for further quantitative research on pulse color and shape features.



Power Vision System Based on Knowledge Distillation Technology

HUANG Shanshan, LUO Wang, HAO Yunhe

2025, 0(10):  20-24.  doi:10.3969/j.issn.1006-2475.2025.10.004

Asbtract ( 37 )   PDF (1087KB) ( 45 )  

References | Related Articles | Metrics


Abstract：This paper proposes a power vision system based on knowledge distillation technology， aiming to solve the application challenges of power vision models in resource-constrained environments. By dynamically selecting a large-scale power vision model with high performance as the teacher model and intelligently constructing a lightweight small model as the student model， the knowledge distillation technology is utilized to compare the outputs of the student model and the teacher model. By minimizing the difference between the two models， effective knowledge transfer is achieved. The research results are applied to the power vision field， which constructs a power vision system including power vision datasets， teacher model selection， student model construction， knowledge distillation， optimized training and model evaluation. Then an automated knowledge distillation process is achieved. Experimental results based on the power vision datasets show that the research can significantly reduce model complexity and computational resource consumption while maintaining high recognition accuracy， improving the applicability and efficiency of the power vision system in resource-constrained environments.



Dense Pedestrian Detection Algorithm Based on Improved YOLOv8

DUAN Jingwei1, CHEN Liang1, LI Xue¹, LIU Mengmeng¹, LIU Jinyu²

2025, 0(10):  24-31.  doi:10.3969/j.issn.1006-2475.2025.10.005

Asbtract ( 54 )   PDF (2315KB) ( 63 )  

References | Related Articles | Metrics

Abstract: To address the issues of missed and false detections in dense pedestrian scenarios caused by complex backgrounds， high crowd density， low-light conditions， and partial occlusions， this paper proposes an optimized dense pedestrian detection algorithm based on YOLOv8n. The algorithm replaces the original convolutional blocks in the backbone network with efficient GSConv convolutions， reducing the model’s computational load while maintaining recognition accuracy. Additionally， GSConv convolutions enable the model to run efficiently on standard GPUs. The feature fusion network is replaced with the SlimNeck lightweight feature fusion module， which reduces the number of feature channels， thereby improving the model’s detection precision and speed. An EMA attention mechanism is embedded in the feature extraction network to enhance the model’s ability to capture both global and local information， thereby reducing false and missed detections in dense pedestrian scenarios. The algorithm also incorporates the Repulsion Loss function to better handle overlaps and occlusions among adjacent pedestrians in dense pedestrian detection， reducing interference between targets and optimizing bounding box regression. Training and validation on the CrowdHuman dataset demonstrate that the improved YOLOv8 model yields a 4.5 percentage points increase in mAP over the baseline. Furthermore, the model exhibits superior performance in dense crowds, occlusions, small-object detection, and low-light conditions, thereby offering an efficient and robust solution for dense pedestrian detection.



Natural Environment Tongue Image Segmentation Method Based on Improved Labv3+ Model

LIU Rongcheng1, XIN Guojiang1, ZHANG Yang1, ZHU Lei2

2025, 0(10):  32-36.  doi:10.3969/j.issn.1006-2475.2025.10.006

Asbtract ( 47 )   PDF (2736KB) ( 47 )  

References | Related Articles | Metrics


Abstract: Tongue image segmentation in natural environments poses great challenges due to factors such as lighting and background interference. This paper proposes a DeepLabv3-MAC model based on improved DeepLabv3+ algorithm for segmenting tongue images captured in natural settings. Firstly, the backbone network of the DeepLabv3+ model is replaced with a MobileNetv2 network to reduce model complexity. Secondly, an asymmetric convolutional module is adopted to enhance the convolutional kernel skeleton of convolutional neural network, thereby improving the utilization of convolutional information. Lastly, the introduction of the CBAM attention mechanism not only focuses on the importance of parameters in space and channels， but also enhances the segmentation accuracy of the model. Experimental results demonstrate that, compared to classical tongue image segmentation algorithms, the proposed DeepLabv3-MAC model exhibits superior segmentation performance. Additionally, the model significantly reduces the number of parameters, enabling faster segmentation of tongue images in natural environments and facilitating future deployment on mobile devices.



Hybrid Model for Electricity Price Prediction Based on DWT-SCINet-MDSC

LI Xue, WEI Yan, LI Linjun

2025, 0(10):  37-43.  doi:10.3969/j.issn.1006-2475.2025.10.007

Asbtract ( 62 )   PDF (1842KB) ( 49 )  

References | Related Articles | Metrics

Abstract: Due to the volatile and complex nonlinear characteristics of electricity prices， the prediction accuracy of existing models is often inadequate. To improve prediction accuracy and dig deeper for valuable information within complex nonlinear characteristics， a hybrid model for electricity price prediction based on DWT-SCINet-MDSC is proposed. Firstly， Discrete Wavelet Transform （DWT） is employed by the model to decompose the data into sub-signals at different time scales. This process not only effectively filtered out high-frequency noise but also more significantly reduced data volatility， thereby enhancing the signal-to-noise ratio and rendering the data clearer and more stable. Secondly， multi-scale separable convolutions are utilized to capture rich information across different time scales while effectively minimizing the number of model parameters， thus accelerating the training process. Lastly， to overcome the limitations of manual feature engineering， a feature weighting module is incorporated to adjust the weights of key features， assigning greater importance to critical features for more efficient feature extraction. A simulation experiment was conducted on an electricity price dataset from a region in Australia. The results indicate that， compared with SCINet and other comparative models， the average absolute error is reduced by 23.29%. This demonstrates that the prediction performance of the DWT-SCINet-MDSC hybrid model is significantly improved.



FA-CGRNet: Non-invasive Classification Model for Hyperglycemia Prediction

WANG Lei, ZHAO Kang, YIN Xiuqiang

2025, 0(10):  44-50.  doi:10.3969/j.issn.1006-2475.2025.10.008

Asbtract ( 65 )   PDF (1580KB) ( 43 )  

References | Related Articles | Metrics

 
Abstract: Current blood glucose detection methods are often invasive， causing inconvenience and potential risks. A non-invasive method using wearable devices to collect physiological data and user-input dietary information for real-time high blood glucose prediction is proposed. To improve prediction accuracy， a deep learning-based time series classification model， FA-CGRNet， is developed. Physiological data are preprocessed through denoising and resampling. Statistical features are extracted. A residual convolutional network is designed to extract and fuse features through convolution and residual connections. A feature enhancement module is utilized to calculate feature weights and perform feature selection. Finally， an LSTM model is employed to extract long-term dependency features from time series data. The model is tested on a public dataset from the BIG IDEAs Lab at Duke University. In the field of non-invasive blood glucose detection， the feature extraction method and network model presented in this study demonstrated superior performance compared to existing time series classification models. The model’s ability to distinguish between positive and negative cases is significantly enhanced. Notably， the weighted F1 score is improved by over 6.2%， while the AUC is increased by more than 2.5%. These results underscore the effectiveness of the proposed approach in advancing non-invasive blood glucose monitoring techniques.



InstantMesh: Three-Dimensional Reconstruction Method for Early Gastric#br# Cancer Images

YAO Minjia1, SONG Wenai1, SUN Xue2, WANG Ziyu3, LEI Yi4, WANG Qing4, ZHAO Li5

2025, 0(10):  51-56.  doi:10.3969/j.issn.1006-2475.2025.10.009

Asbtract ( 57 )   PDF (3780KB) ( 41 )  

Related Articles | Metrics


Abstract: In recent years， the incidence of gastric cancer in China has been continuously rising， while the diagnosis rate of early gastric cancer remains relatively low. As an important means for diagnosing early gastric cancer， magnifying endoscopy can observe micro lesions， but traditional diagnostic methods are difficult to quantitatively analyze， which limits its application in clinical practice and poses a great challenge to the treatment and prognosis of patients. In order to assist in the diagnosis of early gastric cancer， improve the survival rate and prognosis of patients， the 3D reconstruction algorithm of magnified gastroscopy images based on deep learning has become a research hotspot. This paper proposes to use the InstantMesh framework， combined with a multi-view diffusion model and a sparse view reconstruction model， and crops images based on the coordinate information of the lesion area segmented in the previous single magnified gastroscopy image， achieves the construction of a high-quality 3D mesh model for single lesion area images. This method not only improves the reconstruction accuracy， reduces noise interference， but also makes the lesion features clearer. Experimental results show that this method is significantly better than the existing state-of-the-art single-view 3D reconstruction algorithms such as Unique3D， TripoSR， Convolutional Reconstruction Model （CRM） and One-2-3-45 in both qualitative and quantitative evaluation of medical image 3D reconstruction. This study aims to provide strong technical support for early diagnosis and treatment of gastric cancer， makes substantial contributions to improve the prevention and treatment of gastric cancer in my country.



Design and Implementation of Power Design Knowledge Graph Service Platform

XU Yan1, BO Guanglin1, ZHANG Peng1, CHU Cheng2, LIU Wenwen2, ZHANG Bo2, XU Na3

2025, 0(10):  57-66.  doi:10.3969/j.issn.1006-2475.2025.10.010

Asbtract ( 53 )   PDF (1990KB) ( 80 )  

References | Related Articles | Metrics


Abstract: With the explosive growth of data in the field of power engineering design， how to efficiently manage and utilize these data has become an urgent problem to be solved. Based on the knowledge graph technology， this paper focuses on the analysis of the characteristics of the data in the field of power design， and combines the specific needs of the professionals in the field of power design， and designs and implements a power design knowledge graph service platform to achieve efficient management， integration and deep mining of power design data. The platform adopts B/S architecture and follows the J2EE standard， and realizes the functions of power design knowledge search， knowledge graph visualization， intelligent question and answer， similar recommendation， and related data analysis. The test results show that the platform can effectively integrate and manage a large number of power design-related data， and can also deeply explore the correlation and value behind the data through knowledge graph technology， which provides strong support and innovation impetus for the research and application in the field of power design， and significantly improves the correlation and practicability of data. The application of the platform provides powerful knowledge support and decision aid for researchers and engineers in the field of power design， improving their efficiency and ability to process and analyze data. 



Pancreas Segmentation Based on Two-stage Network of Multiple Attention Mechanisms

ZHOU Bangyuan1, XIN Guojiang1, LIANG Hao2, DING Changsong1

2025, 0(10):  67-72.  doi:10.3969/j.issn.1006-2475.2025.10.011

Asbtract ( 53 )   PDF (3109KB) ( 42 )  

References | Related Articles | Metrics


Abstract:Pancreas segmentation is of great significance in computer-aided diagnosis of pancreatic diseases. The pancreas is characterized by small size， large individual differences， and blurred margins， so the task of pancreas segmentation is extremely challenging. To solve the above problems, this paper proposes a new network based on two-stage multi-attention mechanism. Firstly， in order to solve the problem of imbalance between the background and the target， this paper uses a two-stage segmentation method to use the coarse segmentation stage to clip out the candidate region as the input of the fine segmentation stage. Secondly， for the problem of large individual differences in the pancreas， the channel attention mechanism block is designed to the decoder， and the Squeeze and Excited Attention Module（SE Module） is also introduced to adapt to different shapes and sizes of pancreas by using its adaptive attention mechanism. Finally， the Convolutional Attention Module （CBAM） is used to strengthen the information transmission between the encoder and the decoder to improve the segmentation accuracy of the model. The proposed method is tested on the NIH dataset， and the results show that the proposed method has good performance and can effectively solve the problem that the pancreas is difficult to segment in abdominal CT images.



Oxygen Content Prediction of Circulating Fluidized Bed Boiler Based on THBA-BiLSTM 

WANG Zhicong1, MA Yihang2, ZHANG Lingxiang3

2025, 0(10):  73-79.  doi:10.3969/j.issn.1006-2475.2025.10.012

Asbtract ( 46 )   PDF (2096KB) ( 49 )  

References | Related Articles | Metrics


Abstract: Oxygen content is an important parameter reflecting the internal combustion of circulating fluidized bed boiler， for the problem that oxygen content is difficult to predict， a soft measurement model is proposed to improve the bidirectional long and short-term memory network. Firstly， the parameters of input variables such as coal feed， air intake and other input variables related to oxygen content are determined by the input-output correlation coefficient method. Secondly， the output oxygen content soft measurement model is established based on the bidirectional long and short-term memory network， and the BiLSTM prediction model is able to learn the past and the future information， which can better capture the global dependency information. Then， Tent chaotic sequences and Cauchy’s mutation strategy are introduced to optimize the honey-badger algorithm’s initial population， local optimization and global optimization abilities； the improved honey badger optimization algorithm is applied to BiLSTM prediction model parameter optimization， which in turn optimizes the hyperparameters of the BiLSTM model and ensures the accuracy of the measurement model. Finally， the proposed model is applied to the actual output prediction of circulating fluidized bed boiler， and MAE， MSE， MAPE， and RMSE are used as the evaluation indexes， and the experimental results show that the error accuracies of the THBA-BiLSTM neural network proposed in this paper are 1.57e-2， 3.5e-4， 4.1e-3， and 1.87e-2， which are significant enhancement effects relative to the other four models.



Improved FA-BP Neural Network Traffic Flow Prediction Algorithm

WANG Yuanrui, JIANG Lingyun

2025, 0(10):  80-88.  doi:10.3969/j.issn.1006-2475.2025.10.013

Asbtract ( 55 )   PDF (3385KB) ( 45 )  

References | Related Articles | Metrics

Abstract: Traffic flow prediction is one of the important technical means to improve efficiency and reduce congestion in intelligent transportation systems. A BP neural network traffic flow prediction method based on improved Firefly Algorithm （FA） and Levenberg Marquardt （LM） algorithm is proposed to address the problems of slow convergence speed and low prediction accuracy in existing traffic flow prediction algorithms. This method utilizes an improved chaotic Firefly Algorithm to optimize the initial weights and thresholds of the BP neural network， and uses the LM algorithm instead of the traditional gradient descent method in the weight update stage to accelerate the convergence process and improve model accuracy. Finally， the LM-FA-BP algorithm is used to predict traffic flow. Based on real complex urban traffic data， multiple fusion models were compared through experiments. The prediction error of our model was significantly reduced compared to other models， with a 33.84% improvement in Mean Absolute Error （MAE） compared to the BP model and a 29.82% improvement compared to the FA-BP model. The model has been tested and implemented on actual roads， with a maximum accuracy of 98% （average absolute percentage error<2.0%）， reaching a high level. The improved LM-FA-BP model has higher accuracy and faster convergence speed in traffic flow prediction. The research results indicate that the model has broad application prospects， especially in intelligent transportation systems where it can effectively improve prediction accuracy.



Method for Predicting Agricultural Product Prices Based on Improved TimesNet

WANG Yinbing1, 2, WANG Xingfen1, 3, LI Libo1, 3

2025, 0(10):  89-95.  doi:10.3969/j.issn.1006-2475.2025.10.014

Asbtract ( 44 )   PDF (4010KB) ( 46 )  

References | Related Articles | Metrics


Abstract: Predicting agricultural product prices plays a key role in stabilizing the agricultural market. However， due to the influence of various factors， agricultural product prices exhibit characteristics such as non-linearity and periodicity， making it difficult to accurately predict. To solve this problem， a new agricultural product price prediction model， EMD-ConvNeXtV2-TimesNet， is proposed. The model introduces two innovations based on the TimesNet model: first， it innovatively incorporates an Empirical Mode Decomposition （EMD） module to decompose the original price series， thereby better capturing the intrinsic structural information of the price series； second， it improves the image feature extraction module of TimesNet to a ConvNeXtV2 Block to more effectively capture the cyclical information of prices. Comparative experiments were conducted on the collected datasets of corn， eggs， soybeans， and peanuts. The experimental results show that compared with the best prediction results of comparison models such as DLinear， Informer， Transformer， FiLM， FEDformer， the Mean Absolute Percentage Error （MAPE） and Mean Absolute Error （MAE） are reduced by 38.902%/38.562%， 33.183%/33.108%， 39.471%/35.178%， and 48.525%/47.806% respectively. The new model has achieved significant accuracy improvements. Ablation experiments further confirmed the complementary role of the EMD module and ConvNeXtV2 Block in the model， which more effectively reduces the price prediction error compared to the original TimesNet.



Perception and Multimodal Collaborative Risk Assessment Method for Coating Operations

XIE Li1, LIU Zhiying2, ZHANG Yu3, QING Chaojin2

2025, 0(10):  96-102.  doi:10.3969/j.issn.1006-2475.2025.10.015

Asbtract ( 56 )   PDF (1879KB) ( 36 )  

References | Related Articles | Metrics

Abstract: To ensure the efficient operation of distribution network lines， achieving safe and reliable live insulation coating is a critical prerequisite. However, various uncertain risk factors may cause potential safety hazards during the process in electrostatic coating operations.  To address the complexity of these risks and meet the demands for real-time safety monitoring， this paper proposes a risk assessment method for coating operations based on perception and multimodal collaboration. Firstly， the Decision Making Trail and Evaluating Laboratory （DEMATEL） method and Interpretative Structural Modeling （ISM） Method are combined with grey theory and the maximum inter-class variance algorithm to develop a comprehensive DEMATEL-ISM evaluation model. This model systematically analyzes multimodal risk indicators associated with coating operations and assesses the impact of each risk factor. Based on the evaluation results， risk membership functions are constructed for key risk indicators to map real-time operational data into risk membership degrees. Finally， a fuzzy integral is employed to establish a multimodal information fusion model， enabling real-time risk assessment throughout the coating process. This work validates the effectiveness of the proposed method through a case study on risk assessment for live-line insulated coating operations. Evaluation results demonstrate that the method accurately identifies and evaluates critical risk factors， providing strong support for real-time safety monitoring and significantly enhancing operational safety.



Feature Selection Method for Recognizing Covert Mining Behavior

HE Zhiyong1, 2, HE Zeyu1, 2 , ZHANG Wei1, 2 , LIU Guoping3

2025, 0(10):  103-109.  doi:10.3969/j.issn.1006-2475.2025.10.016

Asbtract ( 53 )   PDF (2411KB) ( 42 )  

References | Related Articles | Metrics

Abstract: As the Chinese government continues to regulate cryptocurrency mining activities， miners are increasingly concealing their operations through encryption， proxies， and other methods. Existing mining behavior monitoring techniques have lower accuracy when dealing with covert mining， making effective detection difficult. To address this problem， this paper proposes a covert mining behavior identification method based on RF-Voting. First， we collected and compiled a dataset of covert mining traffic and defined three types of covert mining behaviors. In the feature selection module for covert mining， the RF （Random Forest） feature selector interacts with the Voting classifier to select features， effectively identifying important ones. In the behavior matching module， we propose an enhanced Voting classifier with performance-aware selection and adaptive weight assignment. Performance-aware selection allows for screening high-performance base classifiers， while adaptive weight assignment dynamically adjusts the weights of the classifiers. By combining these two methods， we effectively improve the classification performance and stability of the model. Experimental results show that， compared to traditional mining detection methods， the accuracy of this method was increased by up to 6.18 percentage points， and the F1 score was increased by up to 9.35 percentage points， demonstrating that the RF-Voting method provides a more accurate and effective solution for monitoring covert mining behavior.



BBMA-IBS：Blockchain Bilinear Mapping Arbitration IBS User Identity Authentication

YU Feifei1, 2, SHAO Aiping3

2025, 0(10):  110-117.  doi:10.3969/j.issn.1006-2475.2025.10.017

Asbtract ( 33 )   PDF (3104KB) ( 35 )  

References | Related Articles | Metrics


Abstract: In order to reduce the main system access cost of user identity authentication schemes and improve the security and efficiency of system user identity authentication， a user identity authentication scheme based on blockchain bilinear mapping arbitration identity-based signature （IBS） is proposed. Firstly， a user identity authentication architecture is established using the consortium chain， which includes blockchain， registration server， users， and existing systems. Fuzzy extractors are used to use biometric features for identity authentication， and a user identity unified authentication compatibility model is constructed by combining heterogeneous user digital identities on the chain. Secondly， bilinear mapping is introduced to propose an identity based signature algorithm based on identification， which incorporates an arbitration module to extend the security function of the original algorithm and makes it here a revocation function to enhance system security. Finally， based on the blockchain user identity authentication architecture， process design is carried out for the three modules of system user identity registration， cross domain identity authentication， and intra domain identity authentication. The experimental results show that the algorithm proposed in this paper outperforms the selected comparative algorithms in terms of throughput， latency， and resource overhead， indicating that the proposed algorithm can more effectively improve the security and efficiency of system user identity authentication.



Location Data Collection under Local Differential Privacy

LYU Tianci, LI Yanhui, CHENG Mengyuan, ZHAO Yuxin, HUANG Chen

2025, 0(10):  118-126.  doi:10.3969/j.issn.1006-2475.2025.10.018

Asbtract ( 48 )   PDF (2473KB) ( 44 )  

References | Related Articles | Metrics

Abstract: Privacy-preserving location data collection is an important problem that has attracted much research attention. The state-of-the-art for this problem is local differential privacy （LDP）， which has been established as a strong and rigorous privacy scheme for collecting sensitive information from users. However， existing LDP-based solutions are either limited by poor overall result utility or suffer from large deviation between perturbed and true locations. Motivated by this， we propose CFM， an accurate and efficient method for location data collection and analysis. It generates a safe region based on the overall user’s needs， narrowing down the perturbation output domain and thus enhancing data utility. To make the perturbed location as close to the true location as possible， we devise a novel multi-granularity perturbation strategy by exploiting the proximity between locations. The determination of the optimal range with different granularities is challenge， and CFM accomplishes this through analyzing mutual information. Furthermore， the proposed method integrates probability transition matrices with perturbed location distribution information to estimate the true location frequency distribution. Extensive experiments conducted on real-world datasets demonstrate the effectiveness of our proposed method and its advantages over existing solutions.