DCGAN Based Spectrum Sensing Data Enhancement for Behavior Recognition in Self-Organized Communication Network

Kaixin Cheng,Lei Zhu,*,Changhua Yao,Lu Yu,Xinrong Wu,Xiang Zheng,Lei Wang,Fandi Lin

1The College of Communications Engineering,Army Engineering University,Nanjing 210007,China

2School of Electronic and Information Engineering,Nanjing University of Information Science and Technology.Nanjing 210044,China

Abstract:Communication behavior recognition is an issue with increasingly importance in the antiterrorism and national defense area.However,the sensing data obtained in actual environment is often not sufficient to accurately analyze the communication behavior.Traditional means can hardly utilize the scarce and crude spectrum sensing data captured in a real scene.Thus,communication behavior recognition using raw sensing data under smallsample condition has become a new challenge.In this paper,a data enhanced communication behavior recognition(DECBR)scheme is proposed to meet this challenge.Firstly,a preprocessing method is designed to make the raw spectrum data suitable for the proposed scheme.Then,an adaptive convolutional neural network structure is exploited to carry out communication behavior recognition.Moreover,DCGAN is applied to support data enhancement,which realize communication behavior recognition under small-sample condition.Finally,the scheme is verified by experiments under different data size.The results show that the DECBR scheme can greatly improve the accuracy and efficiency of behavior recognition under smallsample condition.

Keywords:spectrum sensing;communication behavior recognition;small-sample data enhancement;selforganized network.

I.INTRODUCTION

Communication behavior recognition can help people figure out more information about communication individuals or further deduce communication intention.It has significant meaning in both attack and defense technology about anti-terrorism activity and national defense area[1].Nowadays,spectrum sensing has been widely used in communication network security related areas[2–4].However,most researches of spectrum sensing are focusing on spectrum prediction[5]and interference mitigation[6].Theoretically,the spectrum data of a communication network contains not only the information itself to be transmitted,but also the behavior information within the communication network.However,analyzing spectrum data to find out the communication behavior is a topic remains to be explored.

In previous studies,when people want to figure out information of a certain communication behavior in network,they mostly want to decipher the specific content of the communication.Most of these methods are based on monitoring[7–9].For the purpose of enhancing information monitoring ability,The authors in[7]proposed a new method of active monitoring through deception relays to change the source transmission rate.Moreover,an active eavesdropping method based on cognitive interference is proposed to realize more effective monitoring of the communication content[8].In order to increase the possibility of successful eavesdropping,the authors in[9]used a monitor to perform cognitive interference in a half-duplex mode,concentrating transmission at a lower transmission rate to facilitate monitoring.However,these methods have high requirements for sensor monitoring equipment with complicated analysis process,and has nothing to do with the communication content which has undergone complex encryption processing.Therefore,how to analyze the communication networks structure and the behavior of communication individuals without knowing the specific content about signals has become a research direction with both theoretical and practical significance.

Methods based on communication rules came into being.In[10],the spectrum monitoring data was simulated based on the wireless communication rules.The communication relationship was analyzed through statistical analysis of the spectrum monitoring data. Based on the existing relationship spectrum analysis method,authors in[11]used the node distribution and the characteristics of the coordinates in the highdimensional space to realize visualization of the network structure.A communication relationship discovery method based on communication rules is proposed in[12].This kind of method can find the communication relationship of network through mining and analyzing the spectrum data based on communication rules.Although there is no need to understand the specific content,such strong prior condition of known communication rules is almost impossible to achieve in the actual environment.How to get rid of this restriction became a new challenge.

To deal with above problem,an improved method of communication relationship mining based on spatial density clustering(DBSCAN)is proposed in[13]under noisy conditions.It provides brand new prospective for analysis of spectrum monitoring signals.Further,a method for mining communication relationships among massive spectrum stations based on association analysis is proposed in[14].This method improved the peak density clustering algorithm to study the distribution law and statistical information of spectrum signals,which is effective on finding the relevance between chaos and missing spectrum signals.On this basis,authors in[15]used the spectrum data which has been feature-extracted to perform cluster analysis from the dimensions of time,frequency,bandwidth,power,etc.,and then comprehensively matched the clustering results to obtain the communication relationships and network structure.

This kind of methods had been free from reliance on prior knowledge and had no need to acquire specific communicating content.However,the methods require a large quantity of feature-extracted data,which is hardly capable to be obtained in actual scene.Moreover,the feature extracting process means inevitable missing and error.Generally,the process of spectrum data monitoring and collecting is easily to be interrupted by interference.Thus,the collected data is usually in a small amount,and in original spectrum data form,which can be not applicable to the existing means.In addition,current researches about communication behaviors mostly focus on communication relationship,that is,to find out whether there is communication and who is communicating.There is less attention to the behavior of communication individuals and communication intentions.Therefore,how to make full use of the limited raw spectrum monitoring data to fulfill the task of individual behavior recognition in the communication network is still an open problem.

Since Generative Adversarial Network[16](GAN)technology was firstly proposed by Goodfellow in 2014,many derived technologies emerged to improve the performance on data generation[17–23].In the meantime,not a few researches enhanced spectrum data using GAN technology[24–29]to relieve the lack of data.While the existing spectrum data generation only concerned about signal characteristics depiction but not communication behaviors.Meanwhile,as artificial intelligence technology rapidly develops,it has been widely used in the spectrum sensing areas from all respects[30–32].

The study in this paper dispense with knowing specific communication content and communication rules.On the condition of limited raw spectrum data,GAN technology will be exploited on raw spectrum sensing data enhancement.So as to realize communication behavior recognition under small-sample condition.The main contributions made by this article are as follows:

·A data enhanced communication behavior recognition scheme is proposed to solve the difficulty under small-sample conditions.

·An adaptive convolutional neural network structure is exploited to accomplish communication behavior recognition to dispense with reliance on prior knowledge in traditional methods.

·A profile of novel discussion on communication behavior scene construction is provided and verify the validity of proposed scheme and methods.

The rest of this paper is organized as follow:In Section II,a self-organized communication network scene combining with practical situation is constructed and a data preprocessing process proposed.In Section III,a CNN structure is exploited to realize communication behavior recognition with sufficient data set.In Section IV,a data enhancement communication behavior recognition scheme is presented and its validity on small-sample condition is verified through simulations.Finally,the conclusion is conducted in Section V.

II.SCENE CONSTRUCTION AND DATA PREPROCESSING

2.1 Self-organized Communication Scene Construction

The communication scene setting in this paper is multi-band,multi-mode and multi-channel selforganized communication network,which has structure as N={n1,n2,...,nk,...,nN},where N represents a communication network with N nodes,nkis the kth node in network.The communication relationship between nodes in the network can be expressed as C={c1,c2,...,ck,...,cn},in which ckmeans the kth communication pair.During emulation,the communication adopts frequency hopping mode,and the channel selection between different communication pairs is different for the frequency hopping table is different.Define two different communication behaviors:Superior and subordinate communication,Peer communication in network.The initial power of each node,the power propagation loss caused by geographic location,and the channel conditions(signal fading model)are variant in this scene.

Definition of communication behavior:1)Superior and subordinate communication:There is an obvious hierarchical relationship between the two communication sides.The superior sends more and receives less,while the subordinate receives more but send less.2)Peer communication:The two sides of communication do not have an obvious hierarchical relationship,and both sides receive and send in a balanced manner.

Figure 1.Schematic diagram of network communication relationship.

Assume a partially available communication network,in which there are several nodes in network with three levels 1,2 and 3,the priority of which decreases respectively.Among them,Level 1 corresponds to the core level node,Level 2 corresponds to 2 converge level nodes,and Level 3 corresponds to 4 receiving level nodes.Their communication relationship are shown as C={c1,c2,...,c9}in Figure 1:

·Different communication pair ckadopt variant hopping table when making frequency choose.

·Channel fading:The channel fading coefficient of communication pair ckat time t is shown in Eq.(1).α and ε(t)represent the path fading factor and the instantaneous fading coefficient at time t respectively.dkdenotes the physical distance between two sides in communication pair ck.

Then the signal of the communication pair ckmonitored by the monitoring device at time t can be expressed as Eq.(2).sk(t)indicates the original signal transmitted by sender.nk(t)expresses the noise at time t in channel,which is the white Gaussian noise.

·Features of different communication behaviors:When communicating between the superior and the subordinate,the subordinate will send a short reply after the superior sending for a long time;When communicating at the same level,the communication duration of both parties is relatively balanced.

Figure 2.Spectral data X12of communication between node n1and n2.

·Complete at least one communication dialogue for both sides in a monitoring period.

After the above simulation process in Algorithm 1,An amount of Num spectral waterfall diagrams of communication behavior between selected nodes in different time periods can be collected.

Figure 2 and Figure 3 separately denotes communication spectral data collected between node n1and n2,node n2and n3.It can be told from the level of nodes,communication between n1and n2and belongs to superior and subordinate communication,while communication between node n2and n3is peer communication.

The spectral waterfall diagrams display the communication behavior information.The horizontal axis represents the monitoring time,the vertical axis represents the frequency of the signal.Color represents the intensity of signal power at that frequency point and moment.The change of color implies the attenuation of signal power with time and frequency point,which presents abstract time-frequency domain data in a visual way.Though it is intuitive to see the situation in time and frequency domain,the spectrum waterfall diagrams of the two types of communication behaviors are very similar.Therefore,it is necessary to consider how to extract features from existing data to maximize the difference between two communication behaviors in visual layer.There is a need to adopt appropriate preprocessing methods to further process the data.

2.2 Data Preprocessing

It can be found that,through the analysis of the original spectrum data,the insufficient feature differentiation in the two kinds of original data is not conducive to further research.Thus,the purpose of data preprocessing is to maximize the communication behavior features while weakening irrelevant interference.

The data preprocessing process of spectrum sensing data is divided into four steps as shown in Figure 4:

All the operation of data preprocessing is aim to capture the feature of communication behaviors so as to further data enhancement and recognition.Understanding from physical meaning,the color of spectrum diagram is signal intensity,while the variation of color reflects interaction of communication behavior.And the acquired texture after truncation is a refelction of communication behavior information.So,the core principle includes two aspects:1)the pattern should include as much texture as possible 2)the difference between two kinds of communication behaviors should be distinct enough(distinguishable by eyes at least).The first aspect make sure sufficient communication information is collected and the second aspect guarantee the communication behavior features play a dominant role in the collected information.

Firstly,the amplitude compressing processing aims to compress the extremum gap of the spectral diagram,which can provide a more luxuriant intersecting sur-face for next operation.Secondly,we select a proper value to truncate the three-dimensional map to obtain a top view map with abundant texture.Then,we select a piece of spectrum data with the highest power in the diagram and keep the distinction between two kinds of communication behaviors.At last,transform the picture to be monochrome for better data enhancement effect,since the color information will bring more interference to generators.

After this preprocessing process,two types of communication behaviors communication data obtained after preprocessing are shown in Figure 5 and Figure 6.It can be found that the difference between two types of preprocessed communication behavior data is more obvious,which provides a lot convenience for further study.

III.CONVOLUTIONAL NEURAL NETWORK FOR COMMUNICATION BEHAVIOR RECOGNITION

Since the research in this article is based on the spectrum data simulated as actually monitored,it is different from the previous methods which use highly characterized data,such as communication start and end time,communication center frequency,bandwidth,power,etc.Therefore,it is difficult to use analysis methods,such as cluster analysis,feature matching for research.For classification problems with unclear features,using deep neural networks is an effec-tive way,especially for image classification problems.The recognition problem of pre-processed communication behavior data in image is achievable by Convolutional Neural Network(CNN)which can realize complex image recognition problems[33].

Figure 3.Spectral data X23of communication between node n2and n3.

Figure 4.The process of spectrum data preprocessing.

Figure 5.Preprossessed spectral data of communication between node n1and n2.

3.1 Convolution Neural Network Structure

When designing the Convolutional Neural Network for communication behavior recognition,3*3 size convolution kernel is used in order to balance the feature extraction performance and the parameter amount of the model.The increase in network depth improves the nonlinearity of the model and ensures the recognition accuracy of the convolutional network,while the small convolution kernel can compress the amount of parameters to the greatest extent[28].Combined with the characteristics of communication behavior data,the network depth,activation function,and hyperparameter values are separately determined through continuous testing.Finally,a four-layer convolution,twolayer fully connected neural network architecture is chosen after numbers of tests,as shown in Figure 7.When the amount of training data is sufficient,it can reach a quite high recognition accuracy rate.

Figure 6.Preprossessed spectral data of communication between node n2and n3.

3.2 Experiment and Analysis

When each category has 1000 samples,set batchsize as 16,adopting Adam optimizer with learning rate=10-3,and set test size as 0.3.It can be seen in Figure 8 that,training loss(train_loss)and testing loss(val_loss)are rapidly decrease as epoch increase.And train_loss comes to a convergence merely after one epoch,while val_loss needs only about 6 epochs.Training accuracy(train_acc)and testing accuracy(val_acc)both cost one epoch to reach around 1.Two factors lead to this kind of ideal effect:proper preprocessing and sufficient data size.

After the preprocessing in previous section reasonably extracts the spectral data characteristics of the communication behavior,which is helpful for further experimental analysis.Data without preprocessing uses the same network structure can hardly achieve any form of classification.When the two kinds of raw data are feeded into CNN,adjusting the structure and hyper-parameters within acceptable limits,the recognition accuracy is always about 0.5.It means the feature in raw data is too hard to extract by a CNN in acceptable volume.This situation denotes that raw data using the same CNN can hardly achieve any form of classification.That is why the data preprocessing process is needed.It can realize feature extraction which makes the recognition of CNN more easily.

Sufficient training samples provide a large enough learning space for the neural network.Under the condition of reasonable hyperparameter settings,highprecision recognition can be achieved in just a few epochs.However,in practice,the amount of data that sensors can monitor is extremely limited, which means that the size of training data that can be provided to the neural network is small.

Observing Figure 9 it can be seen that when size of training set is very small,that is,under small-sample condition,the training set loss quickly converges but it is difficult to converge on the validation set.It leads to a very high recognition accuracy of the training set but the accuracy of the validation set is maintained at about 0.5,which can not achieve effective recognition indeed.The main reason for this failure is that,it is difficult for the neural network to fit the distribution of the entire data with small training data.The smallsample condition lead to over fitting problem.

Enriching the training samples as much as possible,and solving the problem of lacking diversity and insufficient sample size become a key operation.Consider using the Generative Adversarial Network(GAN)technology to enhance the existing smallsample data.

IV.DATA ENHANCED COMMUNICATION BEHAVIOR RECOGNITION SCHEME

Generative Adversarial Networks[16](GAN) firstly proposed by Goodfellow in 2014,which has a basic structure of two antagonistic networks:Generator and discriminator.The goal of generator is to produce a distribution consistent with the original data,while discriminator is to judge as accurately as possible whether a data is a real sample or a sample generated by the generator.When such a game reaches the state of Nash equilibrium,theoretically,the data generated by the generator has approximately the same distribution with the original data.Therefore,GAN becomes a powerful tool for data generation,data enhancement and other applications under insufficient sample conditions.

Figure 7.Structure diagram of CNN.

Figure 8.The loss and recognition accuracy curve under sufficient data condition(each category has 1000 samples).

4.1 Structure of DECBR Scheme

In this section,a data enhanced communication behavior recognition scheme using GAN technology is proposed to realize communication behavior problem in small sample condition.

The data enhanced communication behavior recognition(DECBR)scheme is shown in Figure 10.After the data preprocessing process,the small-sample data is transformed to a suitable form for further study.DCGAN is utilized to enhance processed data,generating more data after convergence.FID(Frechet Inception Distance)is used to test whether generated data is resemble enough to original data.Once approved by FID test,generated data are thrown into training set of CNN with original data.After all,it will be seen in contrast that,the recognition accuracy of communication behavior has largely increased after this scheme under small-sample condition.

Figure 9.The loss and recognition accuracy curve under small-sample condition(each category has 50 samples).

Figure 10.Data enhanced communication behavior recognition scheme.

Network structure of GAN is presented in Figure 11.The loss function of GAN is as Eq.(3):

Figure 11.Generative adversarial network structure.

Figure 12.Generator structure of DCGAN in DECBR scheme.

In this paper,DCGAN is chosen to realize data enhancement.Deep Convolution Generative Adversarial Networks[21](DCGAN)is the combination of GAN and CNN.In order to facilitate the processing of the image,CNN is used to replace the multi-layer perceptron in the traditional GAN to extract features.And some changes are made in details,such as adding batch normalization to accelerate learning and converging.Select Tanh activation function in the output of the generator and the Relu activation function in other locations.

Figure 13.Discriminator structure of DCGAN in DECBR scheme.

Figure 14.The curve of DCGAN loss value with epochs.

In DECBR scheme,communication behavior data between node and after preprocessing is the real data,which is used to train discriminator.Note here that the sample data sent to DCGAN only contains the training set in previous failed CNN training.For the fairness of experiment,it needs to be guaranteed that the training data used for enhancement does not bring in new samples.The discriminator optimization uses RMSProp(root mean square propagation)algorithm,and the joint optimization of generator and discriminator uses Adam algorithm.After testing and tuning,the optimal generator and discriminator network structure and parameters are determined as shown in Figure 12 and Figure 13.Through training,the loss curve of the generation network and the discriminant network with epoch is obtained as Figure 14.

Figure 15.Exhibition of generated data and real data((a)is Generated data G12of node n1and n2by DCGAN,(b)is Generated data G23of node n2and n3by DCGAN).

As it can be found in Figure 15 that the generated data of the two communication behaviors have obvious differences,where a certain degree of distinction exists.And each type of generated data is also relatively diverse,which means DCGAN does not fall into a single mode.Observation by eyes can tell that each category data is quite similar to the each real data.To further illustrate the effectiveness of generated data,adopt FID(Frechet Inception Distance)[34]as an evaluation index,which has good performance in robustness and discrimination accuracy[35].The FID between the original data and the generated data of the two kinds of communication behaviors is measured.

The main idea of FID for measuring the quality of generated samples is to measure the distance of the feature distribution between real data and generated data.The features are extracted through the Inception network,while the Gaussian model is used to build the feature model,and then the distance between the features is calculated.The smaller the FID value is,the better the quality and diversity of the generated samples are.

Table 1.Comparation of FID between the two kinds data and their generated data.

The calculation formula of FID is as Eq.(4):

where μXand μGare respectively the mean value of the original data set and the feature of the generated data set.ΔXand ΔGrepresent the two covariance matrix of features.

4.2 Discussion on Computational Complexity

To take a more global approach to the performance of this scheme,it is necessary to discuss computational complexity about the proposed methods.Computational complexity contains the time and space complexity,and usually time computational complexity represents the operation time of model,while space computational complexity denotes the total amount of parameters in layers and operational process.For the convolutional layers,the time and space computational complexity are shown as Eq.(5)and(6):

K is the size of convolutional kernel.C is the amount of tunnels,and D is the number of layers.

For fully connected layers,the time and space computational complexity are shown as Eq.(7)and(8):

X is the size of the input matrix and C is the amount of tunnels.

According to the formulas,the computational complexity of the DECBR scheme can be summarized as Table 2,Table 3,Table 4.

Table 2.Computational complexity of CNN part in DECBR scheme.

Table 3.Computational complexity of generator part in DECBR scheme.

Table 4.Computational complexity of discriminator part in DEBCBR scheme.

It is clear to see that the computational complexity of DCGAN part in DECBR have higher orders of magnitude than that of CNN part.It is obvious that the DECBR scheme compress data amount for communication behavior recognition as enlarge the computational complexity due to the DCGAN part.Therefore,the data enhancement should be a balance between effect and efficiency but not unlimited expansion.

4.3 Experiment and Analysis

To make the experiment convincing,the original data set size include 50,100,200,300,500,1000,2000,3000 samples.Firstly,the communication behavior recognition is successively completed under different original data size,and the average test accuracy within 20 epochs is calculated.The average test accuracy can reflect recognition effect and efficiency at the same time.Secondly,data enhancement is conducted by DCGAN in DECBR scheme using different size of original data respectively.At last,consecutive experiments are performed under different enhanced data size,the average accuracy is compared with which under different original data size.

Table 5.Accuracy comparison of different original and enhanced data size.

For fairness,all experiments share the same test set.Considering the computational complexity,the scale of training data should be put into consideration,which means the comparison of accuracy under same training data size is meaningful.

As shown in Table 5,the first column is original data size,the second column represents recognition accuracy of CNN when using the original data as training set.The next columns are accuracy when using the original data and enhanced data as training sets.Different enhanced data size means different amount of enhanced data is added into training set.

To clearly illustrate the effect of different original data size and different enhanced data size,the accuracy is drawn in Figure 16,in which the horizontal axis is enhanced data size and vertical axis is the average recognition accuracy.Different colors and symbols stand for different original data size as the legend shows.Observing Figure 16,there are findings through comparison:

Horizontal comparison:

In most cases,the recognition accuracy obviously increases as the enhanced data size increases when the original data is deficient(under 500 in this experiment).

Vertical comparison:

Figure16.Accuracy comparison of different enhanced data size.

The enhanced data can generally reach same even better recognition ability of 10 times of original data size.For example,the enhanced data size 500 of original data size 50 has the same effect of original data size 500.

The results prove two arguments:Firstly,the data enhancement scheme adopted in this article can realize the enhancement of the characteristic distribution of communication network behavior.Sencondly,the proposed Data Enhanced Communication Behavior Recognition Scheme can effectively resolve the difficulty of communication behavior recognition under small-sample conditions,and greatly improve the accuracy and efficiency of behavior recognition.Experiments imply that the scheme can greatly reduce data set(10% of original amount in this experiment environment)when achieving same recognition.

V.CONCLUSION

This paper discussed a problem on communication behavior recognition based on insufficient spectrum sensing data.To solve this problem,a data enhanced communication behavior recognition(DECBR)scheme was proposed,which include a data preprocessing method and an adaptive CNN structure.Besides,DCGAN is applied to realize data enhancement,which is verified by FID test.At last,the effectiveness of the scheme is proved through comparative experiments.The proposed Data Enhanced Communication Behavior Recognition Scheme can effectively resolve the difficulty of communication behavior recognition under small-sample condition,and greatly improve the accuracy and efficiency of behavior recognition.Experiments imply that the scheme can greatly reduce data set(10% of original amount in this experiment environment)when achieving same recognition.

ACKNOWLEDGEMENT

This work is supported by the National Natural Science Foundation of China(No.61971439 and No.61702543),the Natural Science Foundation of the Jiangsu Province of China(No.BK20191329),the China Postdoctoral Science Foundation Project(No.2019T120987),and the Startup Foundation for Introducing Talent of NUIST(No.2020r100).