Computer Engineering & Science

Review of one-shot neural architecture search

DONG Pei-jie, NIU Xin, WEI Zi-mian, CHEN Xue-hui

2023, 45(02): 191-203. doi:

Abstract ( 248 )

PDF (1150KB) ( 317 ) 　　

The rapid development of deep learning is closely related to the innovation of neural network structure. To improve the efficiency of network architecture design, Neural Architecture Search (NAS), an automated network architecture design method, has become a research hotspot in recent years. Earlier neural architecture search algorithms in iterative search usually have to train and evaluate a large number of sampled candidate networks, which brings huge computational overhead. Through transfer learning, the convergence of candidate network can be accelerated, thus improving the efficiency of neural architecture search. One-shot NAS based on weight transfer technique is based on super graph, and weights are shared among sub graphs, which improves the search efficiency, but it also faces challenging problems such as co-adaptation and ranking disorder. Firstly, we introduce the research related to one-shot NAS based on weight-sharing, and then analyze the key technologies from three aspects of sampling strategy, process decoupling and phase, compare and analyze the search effect of typical one-shot neural architecture search algorithms, and provide an outlook on the future research direction.

Research and analysis of OSPF protocol in mimic defense system

ZHU Xu-quan, JIANG Yi-ming, MA Hai-long, BAO Wan-ning, ZHANG Jin

2023, 45(02): 204-214. doi:

Abstract ( 182 )

PDF (2006KB) ( 181 ) 　　

The mimic defense technology in cyberspace is a new active defense technology based on dynamic heterogeneous redundancy. By introducing multiple heterogeneous redundant executants, the generalized robustness is enhanced. By implementing policies or periodic scheduling for multiple executants, the uncertain changes of characteristics are presented externally to enhance security. The security of routing protocol is an important part of network security. OSPF protocol is the most widely deployed and most complex routing protocol in the real network world. The most urgent problem for network devices that supports mimic defense is how to realize the equivalence of OSPF protocol functions among various heterogeneous implementations. Firstly, the design of mimic defense is described scientifically, the architecture of router supporting mimic defense is described in detail, and the processing method of OSPF protocol in the mimic defense architecture is discussed in depth. The OSPF protocol proxy is introduced to realize the equivalence of OSPF protocol functions among various heterogeneous implementations. The feasibility and effectiveness of this method are verified in a router prototype that supports mimic defense. Finally, the security risks of routers under the conditions of two classic OSPF routing attacks are specifically explained and verified by experiments, which effectively improves the ability to deal with OSPF network attacks.

Review on data processing unit

LIU Zhong-pei, Lv Gao-feng, WANG Ji-chang, YANG Xiang-rui

2023, 45(02): 215-227. doi:

Abstract ( 209 )

PDF (1314KB) ( 303 ) 　　

With the increase of network transmission bandwidth, complex infrastructure operations in data centers oc-cupy more and more computing resources, affecting service processing performance. Data Processor Unit (DPU) is a newly developed special purpose processor. The background is to deal with the performance bottleneck caused by the exponential growth of data volume and complexity in data center. The emergence of DPU is a phased symbol of heterogeneous computing. The core problem to be solved is to reduce the total cost of ownership of the entire system and improve the efficiency of the entire computing system based on the infrastructure. That is, the load that is “inefficient in CPU proces- sing and unable to be handled by GPU” is offloaded to DPU. Firstly, the development background of DPU is introduced, and the hardware architecture of DPU is analyzed based on network processing model. Besides, DPU is compared with smart network interface card and network processor. After that, DPU programming model, DPU products and applications in the industry are introduced. Finally, the future research directions of DPU are summarized and prospected.

Simulation and verification design of 56 Gbps high-speed signal transmission system

LI Bao-feng, LI Tie-jun, LIU Yong-hui, MA Ke-fan, LUO Yu-feng, YAO Xin-an

2023, 45(02): 228-236. doi:

Abstract ( 235 )

PDF (2916KB) ( 235 ) 　　

The high-speed signal transmission system of the new generation high-performance computer is realized by 56 Gbps PAM4 signal, and the transmission channel spans multiple PCB boards and multi-level connectors, so the design of signal integrity is facing great challenges. A simulation and verification scheme of 56 Gbps high-speed signal transmission system for the whole channel is proposed. Through plate parameter calibration, connector parameter test, and PCB wiring model extraction, a more practical complex transmission channel model is established, and the whole channel collaborative simulation experiment is carried out. Through the simulation experiment and iterative design, the stable and reliable transmission of 56 Gbps PAM4 high-speed signal is successfully guaranteed.

A dynamic watermark adjustment strategy in Flink cluster

Lv He-xuan, HUANG Shan, Alkam·Zabibul, WU Si-heng, DUAN Xiao-dong,

2023, 45(02): 237-245. doi:

Abstract ( 108 )

PDF (761KB) ( 145 ) 　　

Two of the most important task metrics that measure data-mining performance specific to big data: one is real-time and the other is accuracy. The stream data flows from data generation to message queue and then into Flink through data source for calculation. In this process, due to different network transmission speed and different computing performance of different nodes, the sequence of stream data entering the computing framework and the time sequence of events generated by data will be partially out of order. The traditional watermark mechanism for window-facing operations cannot consider the real-time performance and accuracy of the operation results in the case of streaming data with uncertain out-of-order degree. To solve this problem, a stream data microcluster model is established. Based on the local out-of-order degree of stream data event time, the out-of-order degree of stream data representing the current moment is calculated by the local out-of-order degree algorithm. A dynamic watermark adjustment strategy is designed to adjust the watermark dynamically according to the degree of flow data disorder. Finally, the dynamic watermark adjustment strategy based on event time window is implemented in Apache Flink framework. Experimental results show that the dynamic watermark adjustment strategy based on event time window can effectively consider the accuracy and real-time performance of window operation under the condition of elastic or uncertain chaotic flow data.

Antifungal drug discovery base on transcriptome data of cell response

YANG Hao-yi, CHEN Wei, YAO Ze-huan, TAN Yu-song, LI Fei

2023, 45(02): 246-251. doi:

Abstract ( 110 )

PDF (532KB) ( 138 ) 　　

With the rapid accumulation of high-throughput data in biomedicine, it comes to be possible to break through the traditional drug design system and establish a rapid discovery method for antifungal drugs starting from rich data characteristics of biomedical information. High-throughput omics data are to calculate similar pharmacodynamic relationships between discovered drugs, which is applied into antifungal drugs discovery. Based on the CMAP and LINCS data platforms, we obtain the cell transcriptome data under the action of the compound as characteristic characterizations of cell's drug effects. Then we measure the similarity between the characterizations with GSEA method and WTCS algorithm. After that, we screen potential antifungal drugs by the comprehensive rank of the similarity of the drugs to be screened and known antifungal drugs. Based on the large-scale calculations of existing antifungal drugs, we found that drugs such as prenylamine and iri-notecan are expected to become antifungal drug candidates, and some of them are supported by related studies, which need to be verified by further experiments. This paper applies biological big data to quick drug rational design and provides important calculation methods for rational design of antifungal drug repositioning, and inspires and accelerates the development of existing antifungal drugs.

An efficient multi-key homomorphic encryption scheme based on common key

LI Wen-qing, MA Rui, ZHANG Wen-tao

2023, 45(02): 252-260. doi:

Abstract ( 227 )

PDF (480KB) ( 311 ) 　　

As one of the ideal implementations of secure multi-party computing, multi-key homomorphic encryption has significant advantages in resisting quantum attacks and facilitating the construction of secure multi-party computing solutions. However, the existing BGV-type multi-key homomorphic encryption algorithm has problems such as complex key calculation and large ciphertext size. Therefore, a multi-key homomorphic encryption scheme is constructed using a single-key homomorphic encryption scheme, which encrypts the main operation part with a single-key homomorphic encryption scheme, and uses the existing multi-key homomorphic encryption to complete the common key and common decryption. Theoretical analysis shows that the encryption scheme can reduce the key size, reduce the complexity of homomorphic multiplication, and improve the efficiency of encryption operations.

Multi-dimensional attribute analysis of industrial control system vulnerability

LI Tong-tong, WANG Shi-rui, ZHANG Yao-fang, WANG Bai-ling, WANG Zi-bo, LIU Hong-ri,

2023, 45(02): 261-268. doi:

Abstract ( 97 )

PDF (787KB) ( 181 ) 　　

In order to solve the problem that the industrial control system vulnerability risk assessment is simple and not closely related to the industrial control environment, a multi-dimensional attri- bute analysis method of industrial control system vulnerability is proposed. Firstly, a template for discriminating vulnerability attack effectiveness and risk category attributes is established, and multi- dimensional evaluation indicators for the degree of risk vulnerability are defined. Secondly, an automat- ed prediction model of risk level based on ernieCat is proposed, which uses the fusion features of vulnerability text descriptions and the intrinsic evaluation attributes of vulnerabilities to predict the seriousness level, hazard level and exploitability level of industrial vulnerabilities. Besides, this paper combines device-level critical information of industrial control system with vulnerability-level risk situations, and establishes multi-dimensional quantitative evaluation indicators to quantitatively assess the risk hazard level for industrial control system vulnerabilities. Experimental results show that the ernieCat model is superior for predicting vulnerability risk level.

An identity-based auditable multiple interception signature scheme

HE Qi-zhi, CAO Su-zhen, WANG Cai-fen, LU Yan-fei, FANG Zi-xuan, YAN Jun-jian

2023, 45(02): 269-276. doi:

Abstract ( 101 )

PDF (794KB) ( 160 ) 　　

To solve the problems of malicious user revisions in content extraction signatures and untraceability of signatures after extraction, an auditable extraction signature scheme is proposed under the identity-based cryptosystem. The scheme adopts a generic model of M-tree to realize hier-archical multiple extraction signatures, and achieves auditability of signatures by backtracking the tree structure to achieve the purpose of extractor auditable questioning rights. Under the random oracle model, based on the discrete logarithmic difficulty problem, it is proved to be resistant to existential forgery under the adaptive selection message attack . The analysis of experimental results shows that the proposed scheme has certain computational advantages in the signature and extraction phases and the signature verification phase.

A DDoS attack detection scheme based on Bi-LSTM in SDN

BAI Jian-jing, GU Rui-chun, LIU Qing-he

2023, 45(02): 277-285. doi:

Abstract ( 225 )

PDF (1469KB) ( 212 ) 　　

Aiming at the DDoS attack threat brought by massive access devices in the 5G Internet of Things (IoT) environment, considering the Software Defined Network (SDN) for the applicability of 5G IoT, a DDoS attack detection scheme using Long Short-Term Memory (LSTM) network in SDN environment is proposed, in order to improve the accuracy of DDoS attack detection. Based on the idea of divide-and-conquer algorithm, a lightweight distributed edge computing architecture, called Only Care Myself (OCM), is proposed. A Bi-LSTM based lightweight neural network is deployed on idle edge nodes in IoT to complete the detection task, which increases the flexibility of detection while maintaining the accuracy. The performance index of the proposed scheme was evaluated on the ISCX2012 dataset, and the feasibility of the proposed scheme is verified. Experimental results show that the proposed scheme can accurately detect DDoS attacks and effectively mitigate DDoS attacks.

A time factor based revocable and traceable attribute-based encryption scheme

XU Cheng-zhou, WANG Chen, ZHANG Wen-tao

2023, 45(02): 286-294. doi:

Abstract ( 104 )

PDF (883KB) ( 160 ) 　　

Existing access policies of attribute-based encryption schemes seldom involve the time factor. When users set access policies for their own data, it is impossible to limit the time when the users who access the data have the attributes. It is also a challenging problem in attribute-based encryption to track and revoke a user who leaks the key maliciously, and the existing revocable schemes are too computationally intensive and inefficient. To address these problems, a revocable and traceable attribute-based encryption scheme based on time factor is proposed. In the scheme, the user's access time is marked in the user key, the earliest/latest time of the user's attribute acquisition can be limited in the access policy, and the time of the user's attribute acquisition is verified during decryption, which enriches the access policy of the system and realizes the backward security of the scheme. The decryption phase is managed by the time verification server, and only the user time tag factor needs to be updated when the user attributes are revoked, and only the time factor needs to be deleted when the user is revoked, so as to achieve efficient revocation and forward security of the scheme. Finally, under the assumption of DBDH, the proposed scheme is IND-CPA secure. The performance analysis and experimental results show that the proposed scheme has richer features and higher performance.

Research on image quality evaluation based on color differential perception of human eye

WANG Yang, LONG Hai-yan, JIA Xi-ran,

2023, 45(02): 295-303. doi:

Abstract ( 126 )

PDF (1108KB) ( 176 ) 　　

Aiming at the problem of saliency image quality evaluation, referring to the differential perception of image color by human vision, an algorithm based on siamese neural network for quality evaluation of image color contrast salient regions is proposed. Firstly, according to the color contrast and its semantic information of the image, the significant regions of color contrast are extracted respectively. Secondly, the extracted color regions are inputted into the siamese neural network as sub-images in the form of sample pairs. Finally, the correlation between the subjective and objective image quality evaluation values is calculated. In the experiment, the Inception-ResNet-V2 network with residual structure is used as the basic model. The EMD loss function is introduced to optimize the distance loss. After being filtered through the Softmax layer, the image quality evaluation value is obtained. Tests are carried out on the TID2013 dataset and the test results show that the proposed algorithm has good performance on this dataset.

Facial image inpainting based on partial convolution and multi-scale feature integration

SUN Qi, ZHAI Rui, ZUO Fang, ZHANG Yu-tao,

2023, 45(02): 304-312. doi:

Abstract ( 246 )

PDF (2006KB) ( 209 ) 　　

Aiming at the problems of the local chromatic aberrations, boundary artifacts and detail defects in inpainting facial images with large damaged areas, this paper proposes a facial image inpainting model based on partial convolution and multi-scale feature integration. The model is divided into two components: a multi-scale inpainting network and a discriminator network. The inpainting network achieves feature extraction and integration of facial images by effectively fusing deep and shallow image features through a multi-level feature extraction module and a main branching module. Moreover, a joint loss function consisting of content loss, perceptual loss, style loss, total variance loss and adversarial loss is constructed for training a multi-scale inpainting network and enhancing visual consistency between generated images and real images through mutual adversarial with discriminator networks. Experimental results show that, under different mask ratio conditions, this model can generate images with reasonable texture structure and contextual semantic information, and perform better in qualitative and quantitative comparisons.

A distributed multi-robot patrolling algorithm based on multi-step length

BAI Yao-wen, DU Ya-jiang, LI Zong-gang,

2023, 45(02): 313-320. doi:

Abstract ( 82 )

PDF (1128KB) ( 121 ) 　　

Aiming at the problem that the average idle time increases because most algorithms in multi-robot patrolling only use the information of the adjacent nodes of the visited nodes, a distributed patrolling algorithm based on multi-step length is proposed. Firstly, an undirected graph is used to model the environment, in which two nodes with paths are neighbours to each other, and the importance degree is introduced to describe the importance degree of the region where the nodes are located. Secondly, the utility function based on the individual robot is designed, in which the neighbour information of the neighbours of the visited node is used. The function combines the idle time and importance of adjacent nodes and considers the local average idle time and the number of nodes of neighbours of adjacent nodes, and then guides the movement of individual robots by comparing the utility functions. Finally, experimental results show that the global average idle time of the proposed algorithm is gradually reduced and has good stability when the number of robots increases gradually. Compared with other algorithms, the proposed algorithm is more suitable for the patrolling task when the number of robots is large.

Heterogeneous task assignment of mobile crowdsensing based on a discrete fireworks algorithm

SHEN Xiao-ning, XU Di, SONG Li-yan, YAO Cheng-bin, WANG Yu-fang,

2023, 45(02): 321-331. doi:

Abstract ( 93 )

PDF (712KB) ( 133 ) 　　

A mathematical model of heterogeneous task assignment problem based on mobile crowdsensing is established. This model considers the psychological and behavioral processes of participants, introduces environmental information, participants health status, credibility, measurement time and other factors, and minimizes the total cost of task completion by finding the optimal task allocation scheme. The total task cost includes compensation cost, data loss cost and journey cost. To solve the model, a discrete fireworks algorithm with prediction information is proposed. This algorithm uses the integer coding, and the heuristic information of distance and matched-degree in the model are adopted to design the firework explosion operator. A grouping linear prediction strategy of explosion amplitude and an adaptive competition mechanism of mutation operator are proposed. Experimental results show that, compared with the existing algorithms, the proposed algorithm can find a better assignment scheme for the heterogeneous task assignment problem with mobile crowdsensing.

Engine fault diagnosis technology based on BA-RVM algorithm

CHEN Cai-sen, HU Hai-rong, CHENG Zhi-wei, FANG Lu-lu

2023, 45(02): 332-337. doi:

Abstract ( 135 )

PDF (578KB) ( 164 ) 　　

Engine is the core component of the armored equipment's power system. Aiming at the problems that the traditional detection technology is difficult to accurately and quickly diagnose the engine fault in a large number of faulty equipment, the diagnosis workload is large, and the efficiency is low, an engine fault diagnosis technology based on the BA-RVM algorithm is proposed. By merging the various parameter index data of typical armored equipment engines, the model is trained by using the data of the acquisition parameter index and the engine fault, so that the model can predict the fault type based on the parameters of the engine. In the model training, the bat algorithm (BA) is proposed to adjust the kernel parameter width of the correlation vector machine algorithm (RVM), so as to obtain the prediction model of the RVM optimal parameters. Finally, the experimental verification of 12/200ZL water-cooled exhaust turbocharged diesel engine is carried out. The experimental results show that, compared with BP algorithm and SVM algorithm, BA-RVM algorithm reduces the error rate of fault diagnosis by 66.67% and 62.5%, respectively.

A cross-language sentiment classification model based on emotional semantic confrontation

ZHAO Ya-li , YU Zheng-tao, GUO Jun-jun, GAO Sheng-xiang, XIANG Yan,

2023, 45(02): 338-345. doi:

Abstract ( 132 )

PDF (660KB) ( 196 ) 　　

Traditional cross-language sentiment classification methods based on machine translation are affected by the performance of machine translation, resulting in lower accuracy of sentiment classification in low-resource languages such as Vietnamese. Aiming at the problem of imbalance between source language and target language markup resources, this paper proposes a cross-language sentiment classification model based on sentiment semantic confrontation. Firstly, the sentences and the emotional words in the sentences are spliced, and the spliced sentences are jointly represented by the convolutional neural network, and the emotional semantic representations in the monolingual semantic space are obtained respectively. Secondly, through the confrontation network, the emotional semantic representations of labeled data and unlabeled data are aligned in the bilingual emotional semantic space. Finally, the most significant representations of sentences and emotional words are spliced together to obtain the results of emotional orientation classification. The experimental results based on the Chinese-English public data set and the Chinese-Vietnamese data set we constructed show that, compared with the mainstream methods of cross-language sentiment classification, the proposed method achieves bilingual sentimental semantic alignment, and can effectively improve the accuracy of sentimental orientation analysis of Vietnamese. The proposed method has obvious advantages in different language pairs.

An improved sparrow search algorithm to optimize SVM for outlier detection

TANG Yu, DAI Qi, YANG Meng-yuan, CHEN Li-fang,

2023, 45(02): 346-354. doi:

Abstract ( 158 )

PDF (877KB) ( 154 ) 　　

Support vector machine (SVM) is a common method for outlier detection. However, there are still common problems that it is difficult to quickly and effectively obtain the optimal paramet- ers, resulting in low detection efficiency and poor stability. In view of this, an improved sparrow search algorithm is proposed to optimize the parameters of SVM. Firstly, the traditional sparrow search algorithm is improved by improved refraction reverse learning and variable logarithm spiral. Then, the improved sparrow search algorithm (ISSA) is used to optimize the parameters of SVM. Finally, the optimized SVM is used in the field of outlier detection. The simulation results show that, under the two evaluation indexes of G-mean and F-measure, the optimized SVM is obviously better than the other three classification algorithms, and has better detection efficiency, stability, and generalization ability.

An elite opposition-based golden sine marine predator algorithm

ZHANG Lei, LIU Sheng, GAO Wen-xin, GUO Yu-xin

2023, 45(02): 355-362. doi:

Abstract ( 195 )

PDF (737KB) ( 176 ) 　　

Because the basic ocean predator algorithm has shortcomings such as low solution accuracy and poor stability at runtime, this paper proposes an elite reverse learning golden sine ocean predator algorithm. The elite reverse learning mechanism is added to improve the population quality of marine predators, and the comprehensive exploration scope of the algorithm is effectively expanded. The golden sine strategy is added to improve the way of hunting their prey of marine predators and reduce the search space of marine predators, and the performance of local development of the algorithm is effectively enhanced. The effectiveness of the improved strategy is evaluated by solving 12 benchmark functions and 12 CEC2017 functions. The solution test results show that the two strategies have an excellent effect on improving the performance of the ocean predator algorithm.

Robustness enhancement oriented multi task machine reading comprehension

TAN Hong-ye, XING Qin-jie

2023, 45(02): 363-369. doi:

Abstract ( 98 )

PDF (544KB) ( 158 ) 　　

At present, Machine Reading Comprehension (MRC) has achieved good success. However, many researches show that MRC models still have some problems in the robustness in terms of over-sensitivity and over-stability. In order to solve these problems, a multi-task MRC model oriented to robustness enhancement is proposed to strengthen the model's ability to understand the passage and the problem. Specifically, in the multi-task learning method, answer extraction is the main task, and the judgment of evidence sentences and the classification of question are auxiliary tasks, which realizes information sharing between these tasks. The experimental results on the robustness test sets show that the proposed model's performance has a significant improvement compared with the baseline models.

Current Issue

Author center

Review center

Online journal