ISSN 
2022 Vol. 51, No. 5
2022, 51(5): 642-650.
doi: 10.12178/1001-0548.2022279
Abstract:
In order to solve the problem that convolutional neural network requires higher and higher memory and time efficiency, a new model for digital image classification is proposed in this paper. The model is a quantum convolutional neural network based on strongly entangled parameterized circuits. Firstly, the classical image is preprocessed and qubit-encoded, and the image feature information is prepared as a quantum state, which is used as the input of the quantum convolutional neural network model. The quantum convolutional layer, quantum pooling layer and quantum full connection layer of the model are designed to extract the main feature information efficiently. Finally, the Z-based measurement is performed on the model output, and the image classification is completed according to the expected value. In this work, the experimental data set is MNIST data. The classification accuracy of {0,1} and {2,7} classification reached 100%. The comparison experimental results show that the QCNN model of the three-layer network structure with average pooling downsampling has higher test accuracy.
In order to solve the problem that convolutional neural network requires higher and higher memory and time efficiency, a new model for digital image classification is proposed in this paper. The model is a quantum convolutional neural network based on strongly entangled parameterized circuits. Firstly, the classical image is preprocessed and qubit-encoded, and the image feature information is prepared as a quantum state, which is used as the input of the quantum convolutional neural network model. The quantum convolutional layer, quantum pooling layer and quantum full connection layer of the model are designed to extract the main feature information efficiently. Finally, the Z-based measurement is performed on the model output, and the image classification is completed according to the expected value. In this work, the experimental data set is MNIST data. The classification accuracy of {0,1} and {2,7} classification reached 100%. The comparison experimental results show that the QCNN model of the three-layer network structure with average pooling downsampling has higher test accuracy.
2022, 51(5): 651-659.
doi: 10.12178/1001-0548.2022268
Abstract:
Anaerobic digestion is a promising technology in the production of renewable energy, in which biogas is the biological energy generated by anaerobic digestion of organic waste. It is necessary to predict and control biogas yield from anaerobic digestion. A multi-quantum filter quantum convolutional neural network with short-term memory is designed. The designed network utilizes the parameterized variational quantum circuit to accept the data 'time window' so as to simulate the short-term memory, and discards too much circuit iteration and parameter number of the multi-quantum filter to make the filter more expressive. In the quantum circuit framework, the optimal convolution and pooling layer circuits are designed to better extract the hidden states in the feature factors. At the same time, the waste management data are strictly preprocessed, and the trend and seasonality in the characteristics are removed by exponential smoothing. The accuracy of the proposed algorithm reaches 83.30%, which is 8% higher than that of the convolutional neural network model (CNN). The root mean square error (RMSE) and mean absolute error (MAE) values are also better than those of artificial neural network (ANN), K-nearest neighbor (KNN) and CNN classical models.
Anaerobic digestion is a promising technology in the production of renewable energy, in which biogas is the biological energy generated by anaerobic digestion of organic waste. It is necessary to predict and control biogas yield from anaerobic digestion. A multi-quantum filter quantum convolutional neural network with short-term memory is designed. The designed network utilizes the parameterized variational quantum circuit to accept the data 'time window' so as to simulate the short-term memory, and discards too much circuit iteration and parameter number of the multi-quantum filter to make the filter more expressive. In the quantum circuit framework, the optimal convolution and pooling layer circuits are designed to better extract the hidden states in the feature factors. At the same time, the waste management data are strictly preprocessed, and the trend and seasonality in the characteristics are removed by exponential smoothing. The accuracy of the proposed algorithm reaches 83.30%, which is 8% higher than that of the convolutional neural network model (CNN). The root mean square error (RMSE) and mean absolute error (MAE) values are also better than those of artificial neural network (ANN), K-nearest neighbor (KNN) and CNN classical models.
2022, 51(5): 661-667.
doi: 10.12178/1001-0548.2022044
Abstract:
Oral squamous cell carcinoma (OSCC) is a malignant tumor idiopathic in the oral cavity, which is the most malignant and most harmful tumor of head and neck. The occurrence and development of oral squamous cell carcinoma is very heterogeneous, caused by the interaction between heredity and environment, including personalized lifestyles. To understand the complex mechanisms of OSCC at the systems biological level, there is an urgent need for a OSCC genetic variation database. In this research, the OSCC genetic variation database (GVDoscc) was established by both natural language processing and manually curating of data from PubMed. GVDoscc contains 1020 genetic variations and 608 clinical sample records extracted from 334 original articles, providing a reliable and accurate open-source for OSCC associated genetic variation pattern identification and molecular mechanism investigation.
Oral squamous cell carcinoma (OSCC) is a malignant tumor idiopathic in the oral cavity, which is the most malignant and most harmful tumor of head and neck. The occurrence and development of oral squamous cell carcinoma is very heterogeneous, caused by the interaction between heredity and environment, including personalized lifestyles. To understand the complex mechanisms of OSCC at the systems biological level, there is an urgent need for a OSCC genetic variation database. In this research, the OSCC genetic variation database (GVDoscc) was established by both natural language processing and manually curating of data from PubMed. GVDoscc contains 1020 genetic variations and 608 clinical sample records extracted from 334 original articles, providing a reliable and accurate open-source for OSCC associated genetic variation pattern identification and molecular mechanism investigation.
2022, 51(5): 668-674.
doi: 10.12178/1001-0548.2022138
Abstract:
Alternative splicing is a key process of gene expression regulation. The histone modifications are closely correlated with alternative splicing, however, its regulatory mechanisms remain unclear. In order to reveal the relationships between histone modifications and alternative splicing, by analyzing the RNA-Seq data, the exon skipping events from human embryonic lung fibroblast cell line were obtained. Subsequently, the relationships among different kinds of histone modifications were analyzed for both excluded and included exons in exon skipping event. Finally, the Bayesian network was built to deduce the casual relationships among histone modifications in exon skipping event.
Alternative splicing is a key process of gene expression regulation. The histone modifications are closely correlated with alternative splicing, however, its regulatory mechanisms remain unclear. In order to reveal the relationships between histone modifications and alternative splicing, by analyzing the RNA-Seq data, the exon skipping events from human embryonic lung fibroblast cell line were obtained. Subsequently, the relationships among different kinds of histone modifications were analyzed for both excluded and included exons in exon skipping event. Finally, the Bayesian network was built to deduce the casual relationships among histone modifications in exon skipping event.
2022, 51(5): 675-680.
doi: 10.12178/1001-0548.2022103
Abstract:
A downlink non-orthogonal multiple access (NOMA) system is studied, in which a base station serves a near user and a far user on the same frequency band simultaneously. Due to physical obstacles or heavy shadowing, there is no direct link from the base station to the far user and the near user acts as a cooperative relay for the far user by adopting the simultaneous wireless information and power transfer (SWIPT) technique. Unlike most existing works, a non-linear model for energy harvesting is adopted to study the system outage performance by considering the non-linear features of the energy harvesting circuits. Specifically, the system outage probability with integral expressions is derived. To characterize the impact of system parameter design on outage probability, the approximated closed-form expression of outage probability can be obtained by further applying the approximation technique, and the optimal design of system parameters is analyzed. Simulation results validate the correctness of the theoretical analysis and the optimal design.
A downlink non-orthogonal multiple access (NOMA) system is studied, in which a base station serves a near user and a far user on the same frequency band simultaneously. Due to physical obstacles or heavy shadowing, there is no direct link from the base station to the far user and the near user acts as a cooperative relay for the far user by adopting the simultaneous wireless information and power transfer (SWIPT) technique. Unlike most existing works, a non-linear model for energy harvesting is adopted to study the system outage performance by considering the non-linear features of the energy harvesting circuits. Specifically, the system outage probability with integral expressions is derived. To characterize the impact of system parameter design on outage probability, the approximated closed-form expression of outage probability can be obtained by further applying the approximation technique, and the optimal design of system parameters is analyzed. Simulation results validate the correctness of the theoretical analysis and the optimal design.
2022, 51(5): 681-687.
doi: 10.12178/1001-0548.2021392
Abstract:
The phase asymmetry of the mixers’ intermodulation (IM) products is studied in this paper based on the second-order memory mechanisms. This study is developed based on the time-varying modulation function (TVMF) and multi-box behavioral mixer (MBBM) model. This phase asymmetry manifests in the phase where the lower and upper third intermodulation (IM3) products have inverse changing trends vs. tone spacing. The research shows that the inverse trends of the phase of the upper and lower IM3 products are caused by baseband modulation and second harmonics modulations. However, the upper and lower IM3 products’ phase inverse trend mechanism created by baseband modulation is different from that created by second harmonics modulations, and thus may be readily distinguished. Finally, the phases of the upper and lower IM3 products of the mixer are simulated based on advanced design system (ADS), and the above analysis is verified with the measured results.
The phase asymmetry of the mixers’ intermodulation (IM) products is studied in this paper based on the second-order memory mechanisms. This study is developed based on the time-varying modulation function (TVMF) and multi-box behavioral mixer (MBBM) model. This phase asymmetry manifests in the phase where the lower and upper third intermodulation (IM3) products have inverse changing trends vs. tone spacing. The research shows that the inverse trends of the phase of the upper and lower IM3 products are caused by baseband modulation and second harmonics modulations. However, the upper and lower IM3 products’ phase inverse trend mechanism created by baseband modulation is different from that created by second harmonics modulations, and thus may be readily distinguished. Finally, the phases of the upper and lower IM3 products of the mixer are simulated based on advanced design system (ADS), and the above analysis is verified with the measured results.
2022, 51(5): 688-693.
doi: 10.12178/1001-0548.2021246
Abstract:
With the development of science and technology, the demands of Internet of vehicle (IoV) and 6G for the fusion of communications and radar (RadCom) technology is gradually increasing. Orthogonal frequency division multiplexing (OFDM) RadCom systems based on sharing continuous-wave have two range processing methods: one based on the periodic autocorrelation function (PACF) and the other based on the frequency domain element level division. In range processing, these two methods have different effects on the received noise, resulting in the difference of radar performance. By analyzing the equivalent noise amplitude amplification factor and relevant sidelobe based on PACF and frequency domain element level division, this paper introduces the calculation method of critical signal-to-noise ratio (SNR) of these two range processing methods. Finally, the effectiveness of the proposed critical SNR calculation method is verified by the simulation evaluation of radar detection performance in the IoV scenario.
With the development of science and technology, the demands of Internet of vehicle (IoV) and 6G for the fusion of communications and radar (RadCom) technology is gradually increasing. Orthogonal frequency division multiplexing (OFDM) RadCom systems based on sharing continuous-wave have two range processing methods: one based on the periodic autocorrelation function (PACF) and the other based on the frequency domain element level division. In range processing, these two methods have different effects on the received noise, resulting in the difference of radar performance. By analyzing the equivalent noise amplitude amplification factor and relevant sidelobe based on PACF and frequency domain element level division, this paper introduces the calculation method of critical signal-to-noise ratio (SNR) of these two range processing methods. Finally, the effectiveness of the proposed critical SNR calculation method is verified by the simulation evaluation of radar detection performance in the IoV scenario.
2022, 51(5): 694-701.
doi: 10.12178/1001-0548.2021249
Abstract:
Aiming at the challenge of real-time implementation and high classification accuracy for hearing aids, a background noise classification algorithm based on the LightGBM ensemble learning is proposed to reduce the computational time in the process. A newly proposed band spectral correlation feature concatenated with the band spectral entropy feature is also presented. This new acoustic feature is formed to improve the noise classification accuracy. Binaural differential signal is used to extract the band spectral features for reducing memory occupation and offline training workload, so as to improve the computational efficiency. Six commonly encountered noise environments of quiet indoors, traffic, wind turbulence, music, cocktail party and vehicle noise from hearing aid research dataset for acoustic environment recognition are considered. The experimental results show that our proposed algorithm significantly improves the performance of the background noise classification in real-time implementation and the accuracy compared with the algorithms based on the random forest model and band features.
Aiming at the challenge of real-time implementation and high classification accuracy for hearing aids, a background noise classification algorithm based on the LightGBM ensemble learning is proposed to reduce the computational time in the process. A newly proposed band spectral correlation feature concatenated with the band spectral entropy feature is also presented. This new acoustic feature is formed to improve the noise classification accuracy. Binaural differential signal is used to extract the band spectral features for reducing memory occupation and offline training workload, so as to improve the computational efficiency. Six commonly encountered noise environments of quiet indoors, traffic, wind turbulence, music, cocktail party and vehicle noise from hearing aid research dataset for acoustic environment recognition are considered. The experimental results show that our proposed algorithm significantly improves the performance of the background noise classification in real-time implementation and the accuracy compared with the algorithms based on the random forest model and band features.
2022, 51(5): 702-708.
doi: 10.12178/1001-0548.2022014
Abstract:
This article explores a photonic column array slow-wave structure (SWS). A 0.28-THz sheet beam backward-wave oscillator (BWO) was designed and simulated by calculating the dispersion, field distribution and particle simulation. When the cathode current density is only 10 A/cm2 (the minimum is less than 6 A/cm2), the voltage is 12.5 kV and the magnetic field is 0.5 T, the structure interacts with the sheet beam by immersion and the output power is 435 mW. On the basis of previous work, the SWS was fabricated using Lithography-Galvanoformung-Abformung (LIGA) fabrication technology. The results show that the column array structure can effectively improve the interaction efficiency and reduce the starting current density; and effectively improve the lifespan of terahertz (THz) vacuum electron device (VED) cathodes, which is a viable means of increasing the performance of THz vacuum radiation sources.
This article explores a photonic column array slow-wave structure (SWS). A 0.28-THz sheet beam backward-wave oscillator (BWO) was designed and simulated by calculating the dispersion, field distribution and particle simulation. When the cathode current density is only 10 A/cm2 (the minimum is less than 6 A/cm2), the voltage is 12.5 kV and the magnetic field is 0.5 T, the structure interacts with the sheet beam by immersion and the output power is 435 mW. On the basis of previous work, the SWS was fabricated using Lithography-Galvanoformung-Abformung (LIGA) fabrication technology. The results show that the column array structure can effectively improve the interaction efficiency and reduce the starting current density; and effectively improve the lifespan of terahertz (THz) vacuum electron device (VED) cathodes, which is a viable means of increasing the performance of THz vacuum radiation sources.
2022, 51(5): 709-714.
doi: 10.12178/1001-0548.2021262
Abstract:
In this study, the electromagnetic pulse (EMP) signals were measured at the Shenguang II Upgrade Range and used as the input source to establish a parallel cable model to reveal the crosstalk phenomenon generated by the EMP coupled cables. The law of crosstalk induced by varying cable distances and cable lengths were investigated, and the influence of cable length on the frequency of crosstalk signals was also unraveled. Meanwhile, the attenuation of the cable to the crosstalk signal after the copper foil shielding is discussed. The results indicate that increasing the distance between parallel cables is beneficial for suppression of crosstalk, while increasing the length of parallel cables reduce the anti-interference ability and lower resonance frequency. Therefore, proper shielding can effectively suppress signal crosstalk and provide experimental and theoretical support for implementing various physical diagnostics safely and accurately.
In this study, the electromagnetic pulse (EMP) signals were measured at the Shenguang II Upgrade Range and used as the input source to establish a parallel cable model to reveal the crosstalk phenomenon generated by the EMP coupled cables. The law of crosstalk induced by varying cable distances and cable lengths were investigated, and the influence of cable length on the frequency of crosstalk signals was also unraveled. Meanwhile, the attenuation of the cable to the crosstalk signal after the copper foil shielding is discussed. The results indicate that increasing the distance between parallel cables is beneficial for suppression of crosstalk, while increasing the length of parallel cables reduce the anti-interference ability and lower resonance frequency. Therefore, proper shielding can effectively suppress signal crosstalk and provide experimental and theoretical support for implementing various physical diagnostics safely and accurately.
2022, 51(5): 715-720.
doi: 10.12178/1001-0548.2021356
Abstract:
According to the theoretical analysis and magnetic field simulation results, an inductively coupled plasma generator for black barrier communication experiment is designed and fabricated. The device can produce a plasma thin layer with the typical characteristics of black barrier plasma sheath with a hemispherical shape, thickness less than 5 cm and plasma density of 1×1018 m−3. Using the method of plasma emission spectrum diagnosis, the electron excitation temperature and electron number density of the plasma produced by the plasma generator are studied, the influence law of coil power on the plasma characteristics is analyzed, and the mechanism is explored, which has a certain reference for the further optimization of the plasma generator and even the radiation control system of electrically small antenna.
According to the theoretical analysis and magnetic field simulation results, an inductively coupled plasma generator for black barrier communication experiment is designed and fabricated. The device can produce a plasma thin layer with the typical characteristics of black barrier plasma sheath with a hemispherical shape, thickness less than 5 cm and plasma density of 1×1018 m−3. Using the method of plasma emission spectrum diagnosis, the electron excitation temperature and electron number density of the plasma produced by the plasma generator are studied, the influence law of coil power on the plasma characteristics is analyzed, and the mechanism is explored, which has a certain reference for the further optimization of the plasma generator and even the radiation control system of electrically small antenna.
2022, 51(5): 721-727.
doi: 10.12178/1001-0548.2021396
Abstract:
The neural network model has been widely used in the fields of artificial intelligence, and has achieved great success. However, the current neural network is facing the problem of adversarial examples attack, which is artificially constructed fake data that can cause a neural network to output incorrect results. This paper proposes an adversarial examples generation algorithm based on searching the decision boundary of neural network. Firstly, weusebinary search between two real samples to find aninitialattacking point. And then,we calculate the normal vector of the neural network on the decision boundary surface, in order to find the most sensitive direction of the neural network. Finally, we usethe direction information to iteratively find the adversarialexample closer to the original data point until the adversarial example converges. By applying the proposed algorithm on the public data sets, the experimental results show that the algorithm can generate adversarial examples with smaller adversarial perturbations, and it can be combined with other attack algorithms to achieve a better attack result.
The neural network model has been widely used in the fields of artificial intelligence, and has achieved great success. However, the current neural network is facing the problem of adversarial examples attack, which is artificially constructed fake data that can cause a neural network to output incorrect results. This paper proposes an adversarial examples generation algorithm based on searching the decision boundary of neural network. Firstly, weusebinary search between two real samples to find aninitialattacking point. And then,we calculate the normal vector of the neural network on the decision boundary surface, in order to find the most sensitive direction of the neural network. Finally, we usethe direction information to iteratively find the adversarialexample closer to the original data point until the adversarial example converges. By applying the proposed algorithm on the public data sets, the experimental results show that the algorithm can generate adversarial examples with smaller adversarial perturbations, and it can be combined with other attack algorithms to achieve a better attack result.
2022, 51(5): 728-736.
doi: 10.12178/1001-0548.2021349
Abstract:
Visual multiple object tracking (MOT) has become a hot issue in computer vision and intelligent analysis of video images. In recent years, with the development of deep learning and practical application needs, more and more one-shot MOT algorithms with outstanding performance have been proposed, attracting much attention from researchers. This paper systematically reviews the popular one-shot MOT algorithms. From different construction ideas, the paper summarizes the motivation, framework design, strengths and weaknesses of methods, research trends, etc. Afterwards, we compare the performances of the one-shot MOT algorithms on the public testing set MOT Challenge, and quantitatively analyze the advantages and limitations of different one-shot methods. Finally, some future thoughts, foresight, and key issues that need to be focused on are introduced based on the research status.
Visual multiple object tracking (MOT) has become a hot issue in computer vision and intelligent analysis of video images. In recent years, with the development of deep learning and practical application needs, more and more one-shot MOT algorithms with outstanding performance have been proposed, attracting much attention from researchers. This paper systematically reviews the popular one-shot MOT algorithms. From different construction ideas, the paper summarizes the motivation, framework design, strengths and weaknesses of methods, research trends, etc. Afterwards, we compare the performances of the one-shot MOT algorithms on the public testing set MOT Challenge, and quantitatively analyze the advantages and limitations of different one-shot methods. Finally, some future thoughts, foresight, and key issues that need to be focused on are introduced based on the research status.
2022, 51(5): 737-742.
doi: 10.12178/1001-0548.2021244
Abstract:
In mobile edge computing, time delay and power consumption are key performance indicators, and they are mutually restricted. This paper studies the joint optimization of time delay and power consumption by distributing tasks between edge and terminal. Firstly, the 0-1 integer programming model for joint optimization of energy consumption and time delay is established on the basis of describing the research problem. Secondly, a branch and bound algorithm for task assignment is designed. Finally, the simulation show that the proposed method can effectively reduce the energy consumption and time delay of mobile edge computing.
In mobile edge computing, time delay and power consumption are key performance indicators, and they are mutually restricted. This paper studies the joint optimization of time delay and power consumption by distributing tasks between edge and terminal. Firstly, the 0-1 integer programming model for joint optimization of energy consumption and time delay is established on the basis of describing the research problem. Secondly, a branch and bound algorithm for task assignment is designed. Finally, the simulation show that the proposed method can effectively reduce the energy consumption and time delay of mobile edge computing.
2022, 51(5): 743-753.
doi: 10.12178/1001-0548.2021288
Abstract:
Aiming at the problem that the diversity of sparrow search algorithm is reduced and it is easy to fall into local optimum in the iterative process, a multi strategy improved sparrow search algorithm (RSSA) based on rough data-deduction is proposed. Firstly, the algorithm initializes the population with the idea of low difference sequence to enhance the global search ability of the algorithm and ensure the integrity of rough data reasoning domain. Then, the rough reasoning data theory is introduced, and the relationship between individuals is established by combining fitness and distance, so as to improve the convergence speed and the ability to jump out of the local optimum. Moreover, the over bounded individuals in the iteration are assigned to the value near the boundary instead of the maximum or minimum value of the boundary at the same time, which ensures the diversity of the population and improves the convergence speed of the algorithm. Compared with the other three algorithms and traditional sparrow search algorithm, the simulation results based on 11 test functions show that RSSA has faster convergence speed, higher accuracy and better effect in the face of multi peak problems.
Aiming at the problem that the diversity of sparrow search algorithm is reduced and it is easy to fall into local optimum in the iterative process, a multi strategy improved sparrow search algorithm (RSSA) based on rough data-deduction is proposed. Firstly, the algorithm initializes the population with the idea of low difference sequence to enhance the global search ability of the algorithm and ensure the integrity of rough data reasoning domain. Then, the rough reasoning data theory is introduced, and the relationship between individuals is established by combining fitness and distance, so as to improve the convergence speed and the ability to jump out of the local optimum. Moreover, the over bounded individuals in the iteration are assigned to the value near the boundary instead of the maximum or minimum value of the boundary at the same time, which ensures the diversity of the population and improves the convergence speed of the algorithm. Compared with the other three algorithms and traditional sparrow search algorithm, the simulation results based on 11 test functions show that RSSA has faster convergence speed, higher accuracy and better effect in the face of multi peak problems.
2022, 51(5): 754-759.
doi: 10.12178/1001-0548.2021136
Abstract:
Feature selection is an essential step in the data preprocessing phase in the field of bioinformatics. Traditional feature selection algorithms ignore the problems of dependency relevance and redundancy between features. This paper proposes a joint feature relevance and redundancy (JFRR) algorithm for feature selection. The algorithm uses mutual information to calculate the redundancy values between features and applies joint mutual information to compute the relevance among the set of selected features, candidate features and class labels. Finally, JFRR is validated with the other six feature selection algorithms on two classifiers using nine different gene datasets with classification accuracy metrics (Precision_micro and F1_micro). The experimental results show that the JFRR method can effectively improve classification accuracy.
Feature selection is an essential step in the data preprocessing phase in the field of bioinformatics. Traditional feature selection algorithms ignore the problems of dependency relevance and redundancy between features. This paper proposes a joint feature relevance and redundancy (JFRR) algorithm for feature selection. The algorithm uses mutual information to calculate the redundancy values between features and applies joint mutual information to compute the relevance among the set of selected features, candidate features and class labels. Finally, JFRR is validated with the other six feature selection algorithms on two classifiers using nine different gene datasets with classification accuracy metrics (Precision_micro and F1_micro). The experimental results show that the JFRR method can effectively improve classification accuracy.
2022, 51(5): 760-765.
doi: 10.12178/1001-0548.2022011
Abstract:
In the network malicious traffic identification task, there is an imbalance between the ratio of the number of malicious traffic samples and the number of normal traffic samples, which leads to poor generalization ability and low recognition accuracy of the trained machine learning model. To solve this problem, this paper proposes a classification method that balances a small number of data classes by using the conditional Wasserstein generative adversarial network (CWGAN-GP) with gradient penalty items based on the visualization of network traffic. This method first uses the network traffic visualization method to segment, fill, and map the original traffic packet capture (PCAP) data into gray-scale images according to the flow as a unit, and then applies the CWGAN-GP method to achieve the balance of the dataset. Finally, in the public dataset USTC-TFC2016 and CICIDS2017, the convolutional neural network (CNN) model is used to classify and test the unbalanced dataset and the balanced dataset. The experimental results show that the balance method using CWGAN-GP is better than the random oversampling, SMOTE, GAN and WGAN balance methods in the three indicators of Precision, Recall, and F1.
In the network malicious traffic identification task, there is an imbalance between the ratio of the number of malicious traffic samples and the number of normal traffic samples, which leads to poor generalization ability and low recognition accuracy of the trained machine learning model. To solve this problem, this paper proposes a classification method that balances a small number of data classes by using the conditional Wasserstein generative adversarial network (CWGAN-GP) with gradient penalty items based on the visualization of network traffic. This method first uses the network traffic visualization method to segment, fill, and map the original traffic packet capture (PCAP) data into gray-scale images according to the flow as a unit, and then applies the CWGAN-GP method to achieve the balance of the dataset. Finally, in the public dataset USTC-TFC2016 and CICIDS2017, the convolutional neural network (CNN) model is used to classify and test the unbalanced dataset and the balanced dataset. The experimental results show that the balance method using CWGAN-GP is better than the random oversampling, SMOTE, GAN and WGAN balance methods in the three indicators of Precision, Recall, and F1.
2022, 51(5): 766-773.
doi: 10.12178/1001-0548.2021403
Abstract:
In malicious portable document format (PDF) detection based on machine learning, the existing features are easy to be confused or escaped. In order to improve the accuracy and robustness of features, this paper studies and improves the feature extraction method based on the existing methods. Combining the content features, structure features and indirect structure features of document object model (DOM) trees, the feature is selected by analyzing the importance of features and finally the malicious PDF detection is realized by using classification algorithm. The structural features are the number of leaf nodes with high-frequency. Content features includes metadata features, byte entropy, stream byte ratio, etc. The improved feature extraction method can avoid the problems of confusion and escape, and improve the accuracy and robustness of features. In the experiments, we extracted and analyzed features from the collected dataset, 58-dim features with high-importance were selected. Then we used LightGBM algorithm to train gradient boosting decision tree. The testing accuracy of this model reaches 99.9%, which is superior to the other methods. In addition, the features of some adversarial samples are simulated, and the detection accuracy is about 99.2%.
In malicious portable document format (PDF) detection based on machine learning, the existing features are easy to be confused or escaped. In order to improve the accuracy and robustness of features, this paper studies and improves the feature extraction method based on the existing methods. Combining the content features, structure features and indirect structure features of document object model (DOM) trees, the feature is selected by analyzing the importance of features and finally the malicious PDF detection is realized by using classification algorithm. The structural features are the number of leaf nodes with high-frequency. Content features includes metadata features, byte entropy, stream byte ratio, etc. The improved feature extraction method can avoid the problems of confusion and escape, and improve the accuracy and robustness of features. In the experiments, we extracted and analyzed features from the collected dataset, 58-dim features with high-importance were selected. Then we used LightGBM algorithm to train gradient boosting decision tree. The testing accuracy of this model reaches 99.9%, which is superior to the other methods. In addition, the features of some adversarial samples are simulated, and the detection accuracy is about 99.2%.
2022, 51(5): 774-785.
doi: 10.12178/1001-0548.2021274
Abstract:
In view of the phenomenon of load transfer and capacity constraints in the connection edges between different networks in reality, a cascading failure model of interdependent networks considering dependent side loads is proposed and applied to the subway bus dependent networks to analyze the load distribution problem caused by the transfer between different transportation networks. In this model, the failure of dependent edge load, failure of dependent nodes and loss of non-maximum connected subgraph are comprehensively considered. By adjusting the attack ratio, this paper analyzes the influence of different load redistribution strategies, different network coupling modes and combination modes, the degree of network dependence and different edge attack modes on the robustness of interdependent networks. The results show that the residual capacity allocation strategy can effectively alleviate the overload failure of dependent edges, and the robustness of interdependent networks composed of small-world networks is better. The robustness of model decreases as the proportion of dependency node increases and increases as the redundancy of dependency increases. Improving the coupling degree and the average degree of sub-networks can effectively improve the invulnerability of interdependent networks. Compared with connected edges, attacking dependent edges has a greater impact on the robustness of networks.
In view of the phenomenon of load transfer and capacity constraints in the connection edges between different networks in reality, a cascading failure model of interdependent networks considering dependent side loads is proposed and applied to the subway bus dependent networks to analyze the load distribution problem caused by the transfer between different transportation networks. In this model, the failure of dependent edge load, failure of dependent nodes and loss of non-maximum connected subgraph are comprehensively considered. By adjusting the attack ratio, this paper analyzes the influence of different load redistribution strategies, different network coupling modes and combination modes, the degree of network dependence and different edge attack modes on the robustness of interdependent networks. The results show that the residual capacity allocation strategy can effectively alleviate the overload failure of dependent edges, and the robustness of interdependent networks composed of small-world networks is better. The robustness of model decreases as the proportion of dependency node increases and increases as the redundancy of dependency increases. Improving the coupling degree and the average degree of sub-networks can effectively improve the invulnerability of interdependent networks. Compared with connected edges, attacking dependent edges has a greater impact on the robustness of networks.
2022, 51(5): 786-799.
doi: 10.12178/1001-0548.2022048
Abstract:
This paper reviews the recent advances in urban modeling and urban computing by fusing big data mining algorithms and system science research methods. The complexity of urban systems is first introduced. The key works on the urban sensing and the evolutionary mechanisms as well as empirical studies of the structure of complex urban systems are summarized. In addition, we reveal possible origins of urban scaling laws across cities and spatial scaling laws within cities, and introduce a unified analytical framework to explain and predict complex urban phenomena across various spatial scales.
This paper reviews the recent advances in urban modeling and urban computing by fusing big data mining algorithms and system science research methods. The complexity of urban systems is first introduced. The key works on the urban sensing and the evolutionary mechanisms as well as empirical studies of the structure of complex urban systems are summarized. In addition, we reveal possible origins of urban scaling laws across cities and spatial scaling laws within cities, and introduce a unified analytical framework to explain and predict complex urban phenomena across various spatial scales.
