Computer and Modernization

Privacy Protection Model for Medical Data Based on HISPAC

YAO Zheng

2022, 0(09): 1-12.

Asbtract ( 228 )

PDF (3711KB) ( 101 )

References | Related Articles | Metrics

The current era is computer time, especially the era of artificial intelligence and big data. The emergence of related industries has led to changes in various industries. As major service industry in China, the medical industry is changing quietly. At the some time the protection technology of medical privacy is developed continuously. With the explosrve growth of data, various types of patient identity information, case information and medical diagnosis information are leaked endless. In order to solve the topic of medical privacy protection, the paper constructs a set of medical privacy protection model, which includes two parts：1) An adaptive neural network privacy risk assessment model is constructed by using Recurrent Neural Network (RNN) and fuzzy reasoning theory, which is used to assign a credit label to the user’s behavior and calculate the privacy risk; 2) Based on the user credit risk value obtained from the model, a set of personal privacy data access control mechanism is established, namely HISPAC (Hospital Information System Privacy Access Control Model). The experiment proves that this mechanism has good privacy protection effect and can effectively solve the problem of medical data privacy leakage.

Chinese Event Detection Based on Multi-lexicon Feature Augmentation

MIAO Zi-jing, MEI Xin

2022, 0(09): 13-18.

Asbtract ( 157 )

PDF (907KB) ( 69 )

References | Related Articles | Metrics

Event detection mainly focuses on event trigger recognition from unstructured text to achieve correct classification of event types. Compared with English, there is no natural separation in Chinese, and word segmentation boundaries need to be determined before lexicon information can be used. In addition, there is a word-trigger mismatch problem in Chinese event detection. According to the characteristics of Chinese and event detection, a Chinese event detection based on multi-lexicon feature augmentation is proposed, which introduces a word collection containing multi-word information into the character-based model through an external dictionary to utilize the semantic information of multiple segmentation results. At the same time, static text word frequency statistics and Chinese word segmentation systems are used to make collaborative decision on the weight of words in the lexical set to obtain more accurate lexical semantics. The experimental results on ACE2005 Chinese dataset show that the proposed method achieves the best performance, which verifies its effectiveness in Chinese event detection.

Employee Behavior Analysis Method Based on Mean Clustering

LI Chun-sheng, FENG Yang-xiao, FU Yu, ZHANG Ke-jia, WU Run-tong

2022, 0(09): 19-24.

Asbtract ( 128 )

PDF (1945KB) ( 75 )

References | Related Articles | Metrics

Aiming at the problem of mining potential behavior rules of enterprise employees under a large amount of heterogeneous data, a behavior analysis method based on mean clustering is proposed. Based on the behavior data of employees in a scientific research institute, a behavior analysis model is established to extract and select behavior characteristics from the access control card data of enterprise employees and professional daily office software data, and the behavior characteristics are analyzed by K-Means cluster analysis. Finally, in terms of work attitude, employees can be roughly divided into diligent, sloppy and ordinary. In terms of job characteristics, employees can be roughly divided into ordinary, professional and management categories. And through the analysis of the clustering results, some hidden behavioral characteristics of the employees are excavated. Through the investigation of relevant personnel on site, combined with the real work nature and position characteristics of employees, it is verified that the data generated by the application of employee behavior in this scenario, combined with the clustering algorithm, can achieve ideal results in the analysis of enterprise employee behavior.

FPGA-based Interactive Control System for Molecular Dynamics Simulation

WANG Xin, WU Jun-hui,

2022, 0(09): 25-31.

Asbtract ( 161 )

PDF (1273KB) ( 66 )

References | Related Articles | Metrics

In molecular dynamics simulation systems, the computation of range-limited forces between molecules requires frequent transfers and large amounts of particle data access. To reduce the computational load on the CPU, FPGAs can be used to accelerate the computation. However, in the FPGA-based molecular dynamics simulation system, the range-limited force calculation module faces huge data transfer pressure and access conflict problems. To address these problems, an interactive control system is designed based on the limited hardware resources on FPGAs. The system consists of a fetch control module and a particle data parsing module. The whole system achieves fast and reliable transmission of particle data from on-chip storage to the range-limited force calculation module through reasonable data arrangement and cooperative work of the two modules. The effectiveness and reliability of the system in processing particle data are further verified by hardware simulation and board-level experiments.

Surface Defect Detection of Automotive Steel Parts Based on Improved YOLOv4

PENG Lu-lu, ZHU Yuan-yuan, JIN Wen-qian, WANG Xiao-mei

2022, 0(09): 32-39.

Asbtract ( 187 )

PDF (2766KB) ( 61 )

References | Related Articles | Metrics

Aiming at the problem of insufficient detection accuracy of YOLOv4 in the data set of surface defects of self-built automobile steel parts, this paper proposes a surface defect detection method of automobile steel parts based on improved YOLOv4 by taking the advantage of deep learning. Firstly, the weighted K-means algorithm is used to determine the initial anchors pre-selection box to enhance the matching accuracy of anchors and feature map size and improve the detection efficiency. Then the SE module is introduced into the residual unit of the YOLOv4 backbone network to increase the weight of useful features and suppress the weight of invalid features to improve the detection accuracy. Finally, the RFB-s module is connected to the 76×76 feature map to enhance the feature extraction ability of small target information. Aiming at the single defect detection problem of self-built data set of surface defects of automobile parts, the experimental results show that the improved model improves the detection accuracy of mAP50 by 4.3 percentage points compared with the original YOLOv4 model, and has a better detection effect on small targets. It shows that the improved algorithm can meet the requirements of detection speed and accuracy for specific steel parts surface defect detection, and effectively solve the practical problems. Aiming at COCO data set multi-classification problem, the mAP50 value of the improved model is 0.2 percentage points higher than that of the original YOLOv4, and the FPS value reaches 20, which indicates that the improved algorithm can be migrated to other data sets, and the generalization of the algorithm is verified.

Entity Recognition Method on EMR Based on Multi-task Learning

YU Peng, CHEN Yu-feng, XU Jin-an, ZHANG Yu-jie

2022, 0(09): 40-50.

Asbtract ( 136 )

PDF (1633KB) ( 80 )

References | Related Articles | Metrics

Named entity recognition of Chinese EMR is the difficulty in medical information extraction. This paper proposes a multi-task learning mechanism to recognize entity which jointly entity recognition and word segmentation training. The private layers based on Bi-LSTM are used to extract private features, the attention network is used as the shared layer and the general feature enhancement mechanism is added to filter the gobal information, which reduces the risk of over-fitting and enhanced the model generalization ability. Moreover, the balanced oversampling method is proposed to augment EMR dataset, which effectively solves the problem caused by the huge discrepancy in EMR entity types. The CCKS2017/CCKS2020 EMR entity recognition dataset and medicine word segmentation dataset are used for joint learning. The experimental results show that the overall performance in EMR entity recognition is significantly improved, and the word segmentation benchmark in medicine dataset is also raised by 3 percent points in F1 value. The detailed analysis show that the proposed model can effectively correct the entity chunking errors caused by irregular writing style, unstructured text or professional nouns/terms in EMR dataset.

Survey of Model Pruning Algorithms

LI Yi, WEI Jian-guo, LIU Guan-wei

2022, 0(09): 51-59.

Asbtract ( 508 )

PDF (1096KB) ( 151 )

References | Related Articles | Metrics

The model pruning algorithms apply different standards or methods to prune the redundant neurons in the deep neural network, which can compress the model to the maximum extent without losing the accuracy of the model, so as to reduce the storage and improve the speed. Firstly, the research status of model pruning algorithm and the main research direction are summarized and classified. The main research areas of model pruning include the granularity of pruning, the method to evaluate the importance of pruning elements, the sparsity of pruning, the theoretical foundation of model pruning, pruning for different tasks and so on. Then, the recent representative pruning algorithms are described in detail as well. Finally, the future research direction in this field is brought forward.

Detection and Location Method for Hub Weld Based on Retinanet

LI Xin, REN De-jun, REN Qiu-lin, CAO Lin-jie, YAN Zong-yi

2022, 0(09): 60-67.

Asbtract ( 127 )

PDF (2927KB) ( 62 )

References | Related Articles | Metrics

This paper proposes a real-time detection and positioning system for hub weld based on deep learning method, designs a visual inspection hardware platform for hub weld, describes the principle of the multi-specification hub weld detection and location algorithm, describes the principle of the object detection algorithm Retinanet based on convolutional neural network and the object detection algorithm CoTNet based on Transformer architecture, optimizes Cot structure and proposes Cotx structure, so that easily replaces the general convolution layer in convolution neural network. Under the Pytorch framework, this paper simplifies the Retinanet network, and optimizes the detection performance of Retinanet network on the hub weld dataset through the fusion and comparison experiment of Cotx structure and Retinanet network. Experimental results show that better detection effets can be obtained by replacing the last few feature extraction layers of Retinanet with Cotx structure. At the production site, the online real-time detection of hub weld is carried out for 30 days, with an average detection accuracy of 99.7% and a single detection time of 7ms, which can meet the requirements of the enterprise production.

DNeStCount: A Data-dependent Encoder-decoder Architecture with Split-attention for Crowd Counting#br#

MENG Xiao-long,

2022, 0(09): 68-77.

Asbtract ( 105 )

PDF (4745KB) ( 53 )

References | Related Articles | Metrics

Crowd count estimation is the linchpin of the crowd management system, which is very important to prevent stampede accident and guide crowd. It has become an increasingly important task and challenging research direction. This paper proposes a data-dependent encoder-decoder architecture with split-attention for crowd counting, called DNeStCount. In order to cope with the challenges of scale variation and perspective distortion of video surveillance, a more dense atrous ratio is applied to the design of the dense atrous spatial pyramid pooling block. In order to improve the accuracy of density map estimation, a learnable and data-dependent upsampling method DUpsampling is applied to the design of the data-dependent feature aggregation. In order to compensate outlier sensitive and untrainable Euclidean loss, Smooth L1 loss is used to the design of loss function. The experiments and analyses on challenging datasets show that DNeStCount is more competitive compared to thoughtful approaches.

Western Blot Image Recognition of Helicobacter Pylori Based on Improved YOLOv5

WANG Meng, ZHANG Hong-xin, LIU Qing-hua, ZHANG Dong

2022, 0(09): 78-84.

Asbtract ( 186 )

PDF (3600KB) ( 56 )

References | Related Articles | Metrics

To solve the problems of low efficiency and slow speed of Helicobacter pylori immunoblot images that heavily rely on visual recognition by physicians, an improved YOLOv5-based Helicobacter pylori immunoblot image detection model is proposed. Firstly, the feature extractor of YOLOv5 is optimized, and DenseNet is used as a new feature extractor to solve the problem of gradient disappearance. Then, by limiting the maximum downsampling multiple, the model is more sensitive to small target detection. Finally, the Swish activation function is introduced to replace the original YOLOv5 LeakyReLU activation function and improve IoU to obtain more accurate boundary information. The detection ability of the improved model is verified in the Helicobacter pylori immunoimprint image data set. The experimental results show that the F1-score of the improved model is as high as 0.93, mAP@0.5 up to 95.4% , mAP@0.5 : 0.95 up to 75.6%, and the detection frames per second is 54 fps, which can meet the clinical detection time limit requirements.

Multi-label Image Classification Method Combined CNN and Interactive Features

WANG Pan-hong, ZHU Chang-ming

2022, 0(09): 85-92.

Asbtract ( 158 )

PDF (1578KB) ( 58 )

References | Related Articles | Metrics

Images exist widely in daily life, and image classification is of great practical significance. Aiming at the problems of low classification accuracy and high computational complexity in current multi-label image classification due to the complexity of the neural network model and the insufficient of extracted image feature information, a multi-label classification method combined CNN and interactive features, namely MLCNN-IF model, is proposed. The model is mainly divided into two steps. Firstly, a lightweight neural network （MLCNN） with only 9 layers is built with reference to the basic structure of traditional CNN, which is used to process image data and extract features. Secondly, based on the features extracted by MLCNN, the combined features of independent features are generated by the interactive feature method, so as to obtain a new and richer feature set. The experimental results show that compared with AlexNet, GoogLeNet and VGG16, the proposed model achieves better classification results on four multi-label image datasets, and its accuracy and precision rate are increased by 9% and 4.8% respectively on average. At the same time, the MLCNN network structure is relatively simpler, which effectively reduces the amount of model parameters and time complexity.

POF Protocol Parser

CHU Su-hong, LIU Lei,

2022, 0(09): 93-98.

Asbtract ( 91 )

PDF (1535KB) ( 62 )

References | Related Articles | Metrics

For the security issues of SDN, the traditional firewalls and antivirus softwares can only prevent unauthorized external intrusions, but have little effect on preventing internal threats such as important information leakage caused by unauthorized modification of switch or controller configurations and flow rules. As the southbound interface of SDN, POF enables the controller to configure and control network behavior. By parsing POF messages, the communication content of SDN can be monitored and internal security problems can be discovered. In this paper, the POF is carefully studied and analyzed, and a protocol parser is designed based on network security audit system, through which the POF message types and key fields can be parsed and identified online, and session logs and operation logs can be generated for storage and display. This helps discover illegal behaviors in time and trace the source of evidence after a cyber security incident occurs. Through experimental tests, the system can achieve at least 30000 connections per second, 460 Mbps throughput, and 530000 packets per second processing performance under the premise of zero packet loss.

Utility-optimized Local Differential Privacy Mechanism for Protecting Location Privacy

FENG Li-gang, ZHU You-wen,

2022, 0(09): 99-105.

Asbtract ( 179 )

PDF (1015KB) ( 71 )

References | Related Articles | Metrics

Mobile devices collect users’ geographic location data to provide personalized services, which will also produce the potential risk of data leakage. The existing geographic location differential privacy protection mechanism treats different geographic location privacy protection levels equally. Utility-optimized local differential privacy (ULDP) considers different levels of privacy protection for data, but it is only applicable to the frequency estimation of category data and has no application in geographic location privacy protection. Considering the geographic location protection scheme under ULDP mechanism, the square mechanism is transformed, and a utility-optimized square mechanism (USM) is proposed. This mechanism meets the local differential privacy for sensitive geographical locations and does not make security requirements for non-sensitive geographical locations to improve the overall utility. Two different real geographic data sets are selected to compare USM with square mechanism under the condition of the same privacy budget. Theoretical analysis and experimental results show that USM has significantly improved in its effectiveness. At the same time, it also looks forward to the possible direction of further optimization of this mechanism.

Virtualization of Secure Access Device Based on Container

JI Yuan, ZHENG Wei-bo, WANG Zi,

2022, 0(09): 106-110.

Asbtract ( 106 )

PDF (1169KB) ( 62 )

References | Related Articles | Metrics

Facing the access requirements of massive power Internet of Things terminals at the information network and Internet boundary in power system, aiming at the problems of uneven resource allocation, poor compatibility, poor scalability and performance bottleneck of various devices at the traditional secure access boundary, a secure access virtualization model based on container is proposed, which adopts DPDK high-performance packet processing framework, mature container cluster management framework, service computing node arrangement and other key technologies completely separate the data plane from the control plane, build an independent data virtualization forwarding plane, and use SR-IOV technology to realize the virtualization of hardware resources and unified scheduling management, and service the security access capability. The security access device cluster based on this model has high performance, high availability, flexible arrangement and strong scalability. The experimental results show that the model can make efficient and rational use of hardware resources and greatly improve the efficiency of power system boundary security access.

Credit Card Fraud Detection Method Based on Improved SMOTE+ENN and XGBoost Algorithm#br#

SUN Dan, SHI Wei-li, RAO Lan-xiang, MENG Sha-sha, GUO Xiao-ming, LI Yi-lun

2022, 0(09): 111-118.

Asbtract ( 195 )

PDF (1685KB) ( 76 )

References | Related Articles | Metrics

With the rapid development of the credit card business in financial institutions, the financial institutions have faced a serious problem in Credit Card Fraud. Aiming at the problem of the unbalanced distribution of the credit card data in the financial institutions, the paper adopts six ways such as the oversampling, the down sampling, the SMOTE+ENN, the SMOTE+Tomeklin, the improved SMOTE+Tomeklin and the improved SMOTE+ENN for processing the unbalanced data. At the same time, the processed six data sets are input into various classification algorithm models for experimental comparison. Then the balance data sets are input into a muilty-classification algorithm model to make experimental comparisons. Finally, a new Credit Card Fraud Detection model combining the improved SMOTE+ENN and XGBoost algorithm is proposed. The empirical results of five evaluation indicators show that the detection method not only improves the discrimination of unbalanced data of Credit Card Fraud, but also improves the accuracy and feasibility of Credit Card Fraud detection.

DGA Domain Name Detection Combining Attention Mechanisms and Parallel Hybrid Network

LIU Li-ting, OU Yu-yi

2022, 0(09): 119-126.

Asbtract ( 133 )

PDF (1278KB) ( 74 )

References | Related Articles | Metrics

Statistical feature-based DGA domain name detection methods relies on complex feature engineering， while the existing end-to-end deep learning methods perform poorly in the multi-classification tasks. To address these problems， a DGA domain name detection method combining attention mechanisms and parallel hybrid networks is proposed. Firstly, deep pyramid convolutional neural networks is introduced to extract deep semantic information of domain names, and DPCNN-SE is proposed by improving DPCNN using the channel attention block called SENet, which can learn inter-channel relationships adaptively and suppress the transmission of useless features. Meanwhile, the self-attention mechanism and the bidirectional long short-term memory network are combined to construct the BiLSTM-SA network to capture the most representative global temporal features in domain name data. Finally, the features extracted by the two networks are fused and fed into the softmax layer to output the classification results. The experimental results show that the method increases the F1-score by 10.30 percentage points and 10.18 percentage points in the multi-classification task of domain name family compared with the single model of CNN and LSTM， respectively; the F1-score increases by 5.97 percentage points and 4.87 percentage points， respectively， compared with the existing hybrid model method Bilbo and BiGRU-MCNN， and has lower computational complexity.

Table of Content