• <tr id="yyy80"></tr>
  • <sup id="yyy80"></sup>
  • <tfoot id="yyy80"><noscript id="yyy80"></noscript></tfoot>
  • 99热精品在线国产_美女午夜性视频免费_国产精品国产高清国产av_av欧美777_自拍偷自拍亚洲精品老妇_亚洲熟女精品中文字幕_www日本黄色视频网_国产精品野战在线观看 ?

    Application of Self-Organizing Feature Map Neural Network Based on K- means Clustering in Network Intrusion Detection

    2019-11-07 03:12:44LingTanChongLiJingmingXiaandJunCao
    Computers Materials&Continua 2019年10期

    Ling Tan,Chong Li,Jingming Xia and Jun Cao

    Abstract:Due to the widespread use of the Internet,customer information is vulnerable to computer systems attack,which brings urgent need for the intrusion detection technology.Recently,network intrusion detection has been one of the most important technologies in network security detection.The accuracy of network intrusion detection has reached higher accuracy so far.However,these methods have very low efficiency in network intrusion detection,even the most popular SOM neural network method.In this paper,an efficient and fast network intrusion detection method was proposed.Firstly,the fundamental of the two different methods are introduced respectively.Then,the selforganizing feature map neural network based on K- means clustering(KSOM)algorithms was presented to improve the efficiency of network intrusion detection.Finally,the NSLKDD is used as network intrusion data set to demonstrate that the KSOM method can significantly reduce the number of clustering iteration than SOM method without substantially affecting the clustering results and the accuracy is much higher than K- means method.The Experimental results show that our method can relatively improve the accuracy of network intrusion and significantly reduce the number of clustering iteration.

    Keywords: K- means clustering,self-organizing feature map neural network,network security,intrusion detection,NSL-KDD data set.

    1 Introduction

    With the status improvement of Internet technologies in modern society,network security now has become a widespread concern.In addition,in pace with the constant upgrades of related industries brought by new technologies such as cloud computing,networking,big data and block-chain,a huge number of distributed users and ubiquitous access to the Internet make the enterprise data boundary no longer exist.Meanwhile,the amount of data is increasing,which leads to a higher value and more security challenges.Moreover,considering the openness of the Internet platform with diverse audiences and connection types,the Internet is vulnerable to external attacks and its security is bound to be greatly threatened[Lin(2015)].In view of these severe issues,intrusion detection technology is particularly important.Network intrusion detection is a research hot spot but also a difficulty in the field of network security and its essence is a clustering problem[Wang,Peng and Luo(2012)].

    As we can see,regardless of labeled supervised learning or unlabeled unsupervised learning,there are different network models and algorithms for data classification problems with the constant development of machine learning.However,due to the high dimension of intrusion feature data of these classification problems in the field of network security as well as small difference between different intrusion categories,many intrusion patterns cannot be accurately clustered.Li et al.[Li,Guan,Zan et al.(2003)]Proposed a network intrusion research based on fuzzy clustering generalized regression neural network(GRNN),in which the method was focus on specified types of intrusion action pattern with the combination of fuzzy clustering and GRNN analysis.Despite of its high classification accuracy and reliability,the method cannot reach the requirements in terms of the diversity of network intrusion.Yi et al.[Yi,Li and Wang(2009)]Proposed a network intrusion detection method based on support vector machine,in which the method still has some problems such as high false alarm,excessive demands data for building up detection models,and low detection performance in small sample conditions without enough training data.By comparing network intrusion detection methods with different network models and algorithms,intrusion detection method based on self-organizing feature map(SOM)neural network is better,which has a shorter training time and a higher detection efficiency[Wang(2007);Wu and Yan(1999);Wan,Wang,Feng et al.(2012)],and it is able to make up deficiencies of the above network models and algorithms,however,the accuracy of network intrusion detection still has room for further improvement,and there are also shortcomings in the application of unsupervised learning.Therefore,according to the advantages ofK-means clustering algorithm to local classification[Kohonen(1995);Jari and Kohonen(1996)]based on the above analysis,an intrusion detection algorithm based onK-means clustering selforganizing feature map neural network(KSOM)network is proposed in this paper.This method even has higher classification accuracy for network intrusion detection and supports unsupervised network anomaly detection.

    This paper is organized in four sections.In the second section,an improved net-work intrusion detection method based onK-means clustering self-organizing feature map neural network(KSON)is comprehensively introduced,as well as introduced respectively theoretical background of the related SOM network model andK-Means clustering algorithm.The third section gives the experimental process and result analysis.In this section,we introduce a new network intrusion detection database.In addition,we propose a new feature extraction method.The conclusion of this paper is drawn in the fourth section.

    2 KSOM theoretical backgrounds

    Due to the high dimensions of the network intrusion feature data,data differences among different intrusion classes are relatively small.Local optimal clustering result cannot be achieved if SOM neural network is simply adopted.And the corresponding learning rate will be low,which makes many intrusion modes unable to be accurately clustered.Note thatK-means clustering algorithm has high clustering performance in dealing with big data sets.And its time complexity is almost linear,whose clustering iteration time is significantly reduced with the advantage of local optimal clustering result.According to the above analysis,the feature of local optimum based onK-means clustering is adopted to improve the SOM neural network to cluster the network intrusion data in this paper.

    2.1 SOM neural network

    SOM neural network was first proposed by the Finnish mathematician Teuvo Kohonen in 1982[Teuvo(1982)].It is an unsupervised competitive learning network that simulates the working characteristics of human brain.By automatically searching the intrinsic and essential attributes of the sample,the parameters and structure of the network are changed with self-organization and self-adaption.It mainly implements the function of mapping similar sample points in the high dimensional space to the adjacent neurons in the network output.For a certain problem,a certain number of neurons are set,and each neuron corresponds to a weight vector.The different classes of data are clustered into different neurons for classification purpose through training[Bukka,Sivarama and Raju(2018)]as shown in Fig.1.

    Figure 1:The network topology of SOM neural network

    It is a feed forward neural network consisting of two layers of neurons,and Fig.1 is divided into the input layer and the competition layer,i.e.,the output layer.The explanation between the two layers is as follows:

    Input layer:The outside information is collected to each neuron of the output layer through the weight vector.

    Output layer:The winning neuron and the neighborhood are obtained through weight competition using competition algorithm.

    The neurons between the two layers are connected by bidirectional connections,and there is no hidden layer in this network.The basic thought of this network model is the corresponding competition opportunities of the competitive layer and the neuron competition to the input mode,and there exists a neuron that eventually becomes the winner of the competition,this winning neuron represents the classification of the input pattern,thus achieving the result of the clustering.The keywords of the SOM mainly include winning neuron,neighborhood and weight adjustment function.

    (1)Winning Neurons(WN)

    For input vectorX(x1,x2,…,xn)and competitive layer(output layer)neuronj,the network weight isWij(i=1,2,…,n;j=1,2,…,m).The distance between the input vector and the weight vector is calculated by Euclidean distance formula Eq.(1),and the minimum value of the distance corresponding to the neuron in the SOM is then obtained.

    The main approach is to promote competition among the output neurons of the network in an attempt to be activated.As a result,there is only one output neuron being activated at each moment,while the statuses of the other neurons are inhibited.This activated neuron is called winning neuron(WN).

    (2)Neighborhood

    For each WN,the input vectorXrepresented by its nearest neuron generally refers to the neuron near WN in SOM,i.e.,the neuron neighborhood(N).

    (3)Weight Adjustment Function

    This function describes the adjustment of WN and N that are changing with the number of iterationstrelative to input vectorX,which can be represented by(2),

    ,wheretis the number of iterations,andηis the learning rate.

    The training process of SOM neural network can be divided into four steps:

    Step 1:Initialize network weightWij,and set the maximum number of iterationsT;

    Step 2:Normalize each input vectorXin NSL-KDD network intrusion detection database,and calculate WN and N;

    Step 3:Update WN,N andWijaccording to the weight adjustment function;

    Step 4:Execute the above steps repeatedly until the number of iterations reaches the given maximum number of iterations.

    2.2 K-means clustering

    TheK-means clustering was proposed Johnson and Wichern in 2002,and then it is probably the most popular methods in the world[Capó,Pérez,Jose et al.(2017)].K-means clustering algorithm is also a most commonly used indirect clustering algorithm based on sample similarity.It belongs to an unsupervised learning method,and the cluster is divided by minimizing the clustering error[Yu,Chu,Wang et al.(2017)].In addition,it is a classic dynamic clustering algorithm on the basis of point wise modification iteration,the key point of which adopt the sum of square errors as criterion function.Its basic thought is to randomly select data objects as the initial center from the data set containing data objects first,then calculate the distance between each data object and each center.According to the nearest neighbor rule,all data objects will be partitioned into the clusters represented by the nearest center.Then the mean of the data objects in each cluster is calculated as the center of each cluster.By comparing the new center and the last center,it can be asserted that if the new center does not change,then the algorithm converges successfully and will output the result,otherwise all the data objects will be re-divided according to the new center until the convergence conditions of the algorithm are satisfied.However,the accuracy of theK-means procedure is very dependent upon the choice of the initial seeds.Thus,this paper presents the efficient strategy to improve theK-means performance which hybrid SOM neural network,the next section will explain in detail carefully.

    The concrete implementation steps are as follows:Given data setX={x1,x2,…xN},xn∈R d,(n=1,2,…N),K-means clustering algorithm will divide the data set intoMdisjoint clusters according to the optimal clustering criteria.The clustering algorithm needs to assign again cluster numberk and initial cluster centermi,and calculate the distance between the input vector of each point to the cluster center through different equations,then a new cluster centercan be obtained,later the training process will repeat until the change of the cluster center is smaller than a given valueε,which is the ending-criterion of the algorithm.

    2.3 KSOM algorithm

    KSOM algorithm proposed in this paper usesK-means clustering algorithm to cluster local data classified by SOM,so that the accuracy of the classification of the whole network intrusion data will reach its best level.The core idea of the KSOM network intrusion detection method proposed in this paper is that the winning neural has reached by SOM to merge byK-means clustering algorithm.Thus,the weight vector of SOM neural network is adopted as the input ofK-means clustering algorithm.Then the input data set is normalized to facilitate model training,and finally sent to the network for training and learning.Besides,in the local clustering process,the network weight of the SOM will be constantly adjusted along with the change of the iterations,and will gradually approximate to the input data.Therefore,we will not solve the SOM cluster center in KSOM algorithm,but only calculate its network weight vector.Then the network weight vector is adopted as the input of theK-means clustering algorithm for local clustering,and finally the result of clustering is achieved.

    The training data set logic chart as shown in Fig.2:

    Figure 2:Entire clustering process using the KSOM method

    The training process of KSOM algorithm can be divided into the following seven steps:

    Step 1:Use the network weights extracted from the training results of the SOM neural network as the input of theK-means clustering,and the input is processed with normalization;

    Step 2:Select k data elements m1,m2,…,mkrandomly as initial cluster center;

    Step 3:Calculate the Euclidean distance between each data point and the cluster center;

    Step 4:Cluster the data in accordance with the criterion that minimum clustering can be obtained when it reaches cluster center;

    Step 5:Adjust the cluster center to regulate Eq.(3),whereNjrepresents the number of data contained in classj,andCjrepresents the data sets of classj;

    Step 7:According to the clustering results of Step 6,cluster locally the original data represented by each network weight value and the algorithm ends.

    Through the above steps,the modification of training algorithm of SOM in this research affects neurons which are not close to the winner neuron on feature space,and the winner neuron pulls more neurons into the same cluster.Further,WN in competitive layer is clustered again by the improved KSOM based onK-means algorithm to ensure that the neurons are not being suppressed in the winning neighborhood(N)of the competitive layer of SOM during the competition.For each iteration,the local WN will lead the adjustment degree of the weight vector change greatly,which makes weight vector become smaller and smaller,and the number of iterations decreases.The final learning rate is significantly improved as well as the clustering accuracy.

    After training,the weight vectors of each neuron in the SOM neural network will separately cluster the input vectors to certain neurons,thus the classification result of different data is obtained.As mentioned above,network intrusion detection is essentially a clustering problem.Therefore,SOM neural network has a certain network intrusion detection effect.Although the intrusion detection method based on SOM neural network has the features of high detection rate,short training time and high versatility,there is still further room for improvement of its detection accuracy,and there are also shortcomings in unsupervised learning applications.Especially the choice of the learning rate makes it have to compromise between the learning rate and stability of the final weight vector,and sometimes the local neurons cannot converge precisely,which makes the initial weight vector of the neuron more distant from the input vector,so that it is unable to win in the competition.Therefore,according to the advantages ofK-means clustering algorithm in local clustering implementation as well as its significant learning rate performance,a KSOM intrusion detection algorithm based on SOM is proposed in this paper.

    3 Experiments procedure and result analysis

    3.1 Data description

    In this paper,NSL-KDD network intrusion detection database is adopted as the training and testing data set[Revathi and Malathi(2013)].Compared with KDD-CUP99 network intrusion detection data set,this data set is provided in CSV format,which has eliminated redundant data in KDD-CUP99 data set,and has overcome the shortcoming that the classifier is biased towards repeated records,and the performance of learning methods is vulnerable to be affected.Moreover,NSL-KDD data set has properly selected the proportion of normal and abnormal data and the volume of test set and training set data is more reasonable,therefore,it is very suitable for model algorithm validation and result evaluation analysis.NSL-KDD network intrusion detection database not only can reflect the structure of network traffic and intrusion conditions,but also can be modified,extended and duplicated.The record of NSL-KDD network intrusion detection database is as shown in Tab.1.

    Table 1:NSL-KDD network intrusion detection database record truncation table

    As shown in Tab.1,NSL-KDD network intrusion detection data set contains five types of data,that is,Normal,Probe,R2L,U2R and Dos,and consists of 125973 training data and 22544 test data sets,each record contains 41 features,such as duration,protocol,service,symbol,source byte and target byte.“attack”vs.“normal” labels as shown in Figs.3,and 4 different types of attacks as shown in Fig.4.

    Figure 3:The distribution of attack and normal

    Figure 4:The distribution of 4 different types of attack

    3.2 Data process

    In this paper,the NSL-KDD data sets are used as training sets and testing sets of our paper,the KSOM utilizes numerical values for the process of data training and testing.Further,the data preprocessing approaches are divided into feature extraction and feature transformation models based on working methodology over the big data processing[Kamei and Kawakami(2018)].Thus,in order to facilitate the next model training,first step in experiments is that turn the character values into numerical values and use the attribute ratio data feature extraction method which ascending feature according to its repeated time for data processing.For example,TCP protocol type have a greater number of repeated time,so we convert TCP protocol type into number 1,etc.as shown in Tab.2.

    Meanwhile,use the forward error correction(FEC)coding technology to encode the classification features into discrete features for the symbolic features of the 41 features,so that the one-dimensional vectors can be transformed into 3 dimensional vectors,such as TCP,UDP and ICMP.Similarly,several other features are mapped into different dimensions of vectors,as shown in Tab.3.

    Table 3:Different symbol feature vector dimensions before and after processing

    Then execute the normalization process of the numerical characteristics,the normalization Eq.(4)is given as follows:

    wherevalis the original value,valnewis the normalized data value,VminandVminare the maximum value and the minimum value of this specific data,respectively.When normalization process is done,each feature will be restricted to an interval of[0,1].After the data set is preprocessed and normalized,network training will start.

    3.3 Experimental process

    In this paper,Python 3.6 environment is adopted to training and testing the NSL-KDD data sets.The implementation of the corresponding pseudo-code is shown in Fig.5.

    Figure 5:KSOM model algorithm pseudo code

    Firstly,data import and initialization processes of the NSL-KDD network intrusion detection data set are implemented as well as normalization.Secondly,the clustering algorithm is executed including the judgment of the end condition.Then the output is transferred to the next layer ofK-means clustering model.Similarly,the data normalization is first implemented,then clustering process starts,and finally the judgment of the end condition is implemented.When both of the two end conditions are satisfied,the clustering result is obtained.

    In this paper,we tested respectively SOM neural network,K-means and KSOM method with NSL-KDD set data.First approach is SOM,the learning rateηwas set to 0.01.Second approach isK-means.For over fitting not happen we set the kind numberskequal 8 and the initial point is chosen as far as possible.The proposed KSOM method the learning rateηwas also set to 0.01.But there is a difference that the initial point was chosen as the weight vectors of each neuron in the SOM neural network.

    3.4 Result analysis

    After repeated training and we have tested three different methods,as shown in Fig.6,the average error of SOM neural network clustering along with the increase of iterations.It can be seen from Fig.6,with the increase of iterations,the average error is gradually decreased,yet the update of the weight vector of SOM neural network in the process of training is not global optimal,thus the sum of the average error tends to be increased slightly,followed by the adjustment it begins to decline,hence it will show a saw tooth downward decline.And although the sum of the average error is small,the number of iterations takes 1600 times before it starts to stabilize.Compare the Fig.7 the average error ofK-means clustering along with the increase of iterations,the number of iterations have a substantial reduction which reached 10 times.Unfortunately this method is at the expensive of increased average error.And the proposed of KSOM method as shown in Fig.8,It can be seen that clustering ends after 80 iterations,the error of which stopped at a certain value,and there is also no saw tooth downward attenuation found.It means that the KSOM method is able to solve the problem that SOM neural network cannot achieve a local optimum,and the number of iterations is reduced to a large extent,which improves the clustering efficiency.Meanwhile,the KSOM method also deals with the problem of lower clustering result which happened in theK-Means method.

    Figure 6:The average error decay process of SOM clustering

    Figure 7:The average error decay process of K- Means clustering

    Figure 8:The average error decay process of KSOM clustering

    The next in this paper,accuracy(A C)is adopted as the criterion to evaluate the network intrusion accuracy.The calculation Eq.(5)is given as below:

    whereTPis the number of correct classified attack behavior samples,TNis the number of correct classified normal behavior samples,FPis the number of wrongly classified normal behavior samples,andFNis the number of wrongly classified attack behavior samples.

    In order to test the effect of the proposed KSOM network intrusion detection method based onK-means,NSL-KDD network intrusion detection data set was also trained and tested on other methods,and a comparison was made on the performances before and after the improvement of SOM neural network clustering method,as shown in Tab.3.

    Table 3:Comparison of clustering methods

    From Tab.3 we can see that the accuracy of the improved KSOM based onK-means algorithm is higher than that of the original counterpart under the same data set.In addition,the KSOM model proposed in this paper not only is the number of iterations better than other algorithms,it also has good results in accuracy.Thus,this paper also compares other existing network intrusion detection algorithms on the accuracy of intrusion detection under the same data set.The comparison results are shown in Tab.4.The result shows that the proposed KSOM based onK-means algorithm has a better performance in network intrusion detection.

    Table 4:Comparison of accuracy of intrusion detection in different method

    4 Conclusion

    To deal with the drawback of self-organizing neural network that it has a large number of iterations in network intrusion detection,a KSOM model was proposed in this paper by utilizing the characteristic thatK-means clustering algorithm has faster clustering speed.The experimental results showed that the iterations of the KSOM method was significantly reduced in comparing with self-organizing neural network model and the accuracy was significantly improved with respect to other methods of network intrusion.

    Meanwhile,our method achieved a higher efficiency relatively when implementing network intrusion detection.Thus,the KSOM method has high performance to provide more reliable data and support for the establishment of network security mechanism.

    亚洲av二区三区四区| 国产中年淑女户外野战色| 床上黄色一级片| 日本与韩国留学比较| 国产精品1区2区在线观看.| 久久久久网色| 成人三级黄色视频| 三级男女做爰猛烈吃奶摸视频| 国产又色又爽无遮挡免| 天堂影院成人在线观看| 晚上一个人看的免费电影| 亚洲最大成人手机在线| 亚洲精品国产成人久久av| 亚洲一级一片aⅴ在线观看| 男的添女的下面高潮视频| 日韩,欧美,国产一区二区三区 | 一个人看视频在线观看www免费| 日韩欧美国产在线观看| 国产精品一二三区在线看| 亚洲综合精品二区| 久久精品久久精品一区二区三区| 亚洲熟妇中文字幕五十中出| 亚洲伊人久久精品综合 | 亚洲av成人精品一二三区| 亚洲精品亚洲一区二区| 亚洲经典国产精华液单| 国产日韩欧美在线精品| 国产精品国产三级国产专区5o | 欧美高清成人免费视频www| 亚洲国产精品国产精品| 欧美一级a爱片免费观看看| 十八禁国产超污无遮挡网站| 如何舔出高潮| 久久人人爽人人片av| 91精品伊人久久大香线蕉| 精品午夜福利在线看| 2022亚洲国产成人精品| 人体艺术视频欧美日本| 97超碰精品成人国产| av线在线观看网站| 日韩 亚洲 欧美在线| 国产在视频线精品| 一二三四中文在线观看免费高清| 亚洲av电影在线观看一区二区三区 | 久久精品久久久久久噜噜老黄 | 色综合站精品国产| 精品久久久久久电影网 | 97热精品久久久久久| av卡一久久| 午夜亚洲福利在线播放| 免费大片18禁| 久久久久久久久大av| 欧美bdsm另类| 哪个播放器可以免费观看大片| 午夜a级毛片| 91aial.com中文字幕在线观看| 日本黄色视频三级网站网址| 一个人看视频在线观看www免费| 联通29元200g的流量卡| 亚洲无线观看免费| 精品久久久久久久人妻蜜臀av| av在线老鸭窝| 高清在线视频一区二区三区 | 麻豆精品久久久久久蜜桃| 卡戴珊不雅视频在线播放| 狂野欧美激情性xxxx在线观看| 男人舔女人下体高潮全视频| 亚洲中文字幕一区二区三区有码在线看| 国产精品.久久久| 黄色欧美视频在线观看| 精品一区二区三区视频在线| 美女内射精品一级片tv| 成人无遮挡网站| 色哟哟·www| 亚洲精品成人久久久久久| 久久热精品热| 小说图片视频综合网站| 成人高潮视频无遮挡免费网站| 一级爰片在线观看| 欧美极品一区二区三区四区| 久久人人爽人人片av| 精品一区二区三区视频在线| 成人性生交大片免费视频hd| 亚洲高清免费不卡视频| 精品久久久久久久末码| 午夜老司机福利剧场| 亚洲av熟女| 最近视频中文字幕2019在线8| 日韩欧美三级三区| 欧美精品国产亚洲| 中文精品一卡2卡3卡4更新| 26uuu在线亚洲综合色| 成人毛片a级毛片在线播放| h日本视频在线播放| 欧美最新免费一区二区三区| 九草在线视频观看| 成年版毛片免费区| 欧美区成人在线视频| 日本黄大片高清| 国产精品1区2区在线观看.| 夫妻性生交免费视频一级片| 精品酒店卫生间| 伦精品一区二区三区| 九草在线视频观看| 国语自产精品视频在线第100页| 黄色配什么色好看| 麻豆成人av视频| 国产成人freesex在线| 麻豆成人av视频| 一个人看视频在线观看www免费| 伦精品一区二区三区| 国产精品国产高清国产av| 亚洲av二区三区四区| 人体艺术视频欧美日本| 国产精品不卡视频一区二区| 日韩欧美国产在线观看| 亚洲精品456在线播放app| a级毛色黄片| 亚洲欧美中文字幕日韩二区| 亚洲国产最新在线播放| 91在线精品国自产拍蜜月| 色哟哟·www| 一级二级三级毛片免费看| 免费观看性生交大片5| 91精品伊人久久大香线蕉| 亚洲熟妇中文字幕五十中出| av又黄又爽大尺度在线免费看 | 我的女老师完整版在线观看| 精品一区二区免费观看| 长腿黑丝高跟| 亚州av有码| 美女黄网站色视频| www.色视频.com| 久久久久久久久大av| 高清日韩中文字幕在线| 久久亚洲精品不卡| 最新中文字幕久久久久| 国产极品天堂在线| 波多野结衣高清无吗| 亚洲精品乱码久久久久久按摩| 国产女主播在线喷水免费视频网站 | 免费观看的影片在线观看| 日本一二三区视频观看| 99在线视频只有这里精品首页| 欧美xxxx黑人xx丫x性爽| av女优亚洲男人天堂| 最近的中文字幕免费完整| 嘟嘟电影网在线观看| 午夜久久久久精精品| 国产精品野战在线观看| 日韩,欧美,国产一区二区三区 | 深爱激情五月婷婷| 18禁在线无遮挡免费观看视频| 性色avwww在线观看| 男人和女人高潮做爰伦理| 国产精品一二三区在线看| 国产一级毛片在线| 久久精品国产鲁丝片午夜精品| 一级毛片aaaaaa免费看小| 国产探花极品一区二区| 亚洲性久久影院| 欧美三级亚洲精品| 人人妻人人澡欧美一区二区| 亚洲综合色惰| 我要看日韩黄色一级片| 欧美成人一区二区免费高清观看| 在线免费十八禁| 建设人人有责人人尽责人人享有的 | 亚洲中文字幕一区二区三区有码在线看| 国产午夜福利久久久久久| 亚洲精品日韩av片在线观看| 国产真实伦视频高清在线观看| 看非洲黑人一级黄片| 久久久成人免费电影| 男女边吃奶边做爰视频| 久久99热这里只频精品6学生 | 秋霞在线观看毛片| 免费在线观看成人毛片| 九九爱精品视频在线观看| 午夜视频国产福利| 午夜福利在线观看免费完整高清在| av免费观看日本| 一边亲一边摸免费视频| 性插视频无遮挡在线免费观看| 中文乱码字字幕精品一区二区三区 | 久久婷婷人人爽人人干人人爱| 午夜福利高清视频| 日韩av不卡免费在线播放| 看十八女毛片水多多多| 两个人的视频大全免费| 1000部很黄的大片| 亚洲精品,欧美精品| 男人舔女人下体高潮全视频| 久久人人爽人人爽人人片va| 69av精品久久久久久| 精品一区二区三区视频在线| 美女高潮的动态| 中文字幕人妻熟人妻熟丝袜美| 久久久精品大字幕| 欧美日本视频| 夫妻性生交免费视频一级片| 一区二区三区高清视频在线| 亚洲真实伦在线观看| 欧美bdsm另类| 亚州av有码| 免费看美女性在线毛片视频| 国模一区二区三区四区视频| 亚洲成人久久爱视频| 成人二区视频| 午夜久久久久精精品| 成人午夜精彩视频在线观看| 伊人久久精品亚洲午夜| 精品国产一区二区三区久久久樱花 | 日本av手机在线免费观看| 18禁在线无遮挡免费观看视频| 国产老妇伦熟女老妇高清| 亚洲不卡免费看| 亚洲av中文字字幕乱码综合| 免费看光身美女| 亚洲精品aⅴ在线观看| 免费观看a级毛片全部| 国产成人福利小说| 国语对白做爰xxxⅹ性视频网站| 亚洲av电影不卡..在线观看| 亚洲性久久影院| 日本免费一区二区三区高清不卡| 日韩欧美精品v在线| 日韩大片免费观看网站 | 三级毛片av免费| 亚洲国产精品国产精品| 久久久久久久久中文| 精品不卡国产一区二区三区| ponron亚洲| 成人国产麻豆网| 亚洲精品国产成人久久av| 99久久无色码亚洲精品果冻| 国产精品久久久久久精品电影小说 | 亚州av有码| 亚洲精品色激情综合| 舔av片在线| 亚洲欧美日韩高清专用| 嘟嘟电影网在线观看| 联通29元200g的流量卡| 亚洲中文字幕一区二区三区有码在线看| 日韩一本色道免费dvd| 七月丁香在线播放| 色播亚洲综合网| 水蜜桃什么品种好| 成年av动漫网址| 麻豆一二三区av精品| 日韩精品青青久久久久久| 欧美性感艳星| 又爽又黄无遮挡网站| 国产精品野战在线观看| 美女脱内裤让男人舔精品视频| 久久国产乱子免费精品| 亚洲最大成人av| 午夜精品在线福利| 成人鲁丝片一二三区免费| 一级毛片aaaaaa免费看小| 永久网站在线| 日韩中字成人| 丰满乱子伦码专区| 18禁裸乳无遮挡免费网站照片| 国产v大片淫在线免费观看| kizo精华| 成人毛片60女人毛片免费| 自拍偷自拍亚洲精品老妇| 亚洲av福利一区| 亚洲av熟女| 午夜激情欧美在线| 国产又黄又爽又无遮挡在线| 中文资源天堂在线| 亚洲国产高清在线一区二区三| 亚洲欧美日韩无卡精品| 五月玫瑰六月丁香| 欧美激情在线99| 国产老妇伦熟女老妇高清| 亚洲av电影不卡..在线观看| 成人午夜精彩视频在线观看| av福利片在线观看| av播播在线观看一区| 国产国拍精品亚洲av在线观看| 91精品伊人久久大香线蕉| 在线免费观看不下载黄p国产| 亚洲美女搞黄在线观看| 久久久久久久久中文| 国产熟女欧美一区二区| 国产午夜精品久久久久久一区二区三区| 亚洲自偷自拍三级| 女人十人毛片免费观看3o分钟| 一级黄片播放器| 久久久精品大字幕| 久久精品91蜜桃| 亚洲怡红院男人天堂| 久久精品久久精品一区二区三区| 成人鲁丝片一二三区免费| 中文资源天堂在线| 小说图片视频综合网站| 两个人的视频大全免费| 国产一区二区三区av在线| 少妇丰满av| 视频中文字幕在线观看| 麻豆成人午夜福利视频| 中文字幕久久专区| 国内精品美女久久久久久| 国产精品av视频在线免费观看| 少妇裸体淫交视频免费看高清| 非洲黑人性xxxx精品又粗又长| 联通29元200g的流量卡| 最近中文字幕2019免费版| 色综合亚洲欧美另类图片| 一区二区三区四区激情视频| 麻豆乱淫一区二区| 久久这里有精品视频免费| 欧美日韩一区二区视频在线观看视频在线 | 久久久久久久久久成人| 日韩欧美 国产精品| 国产精品三级大全| 能在线免费观看的黄片| 青春草视频在线免费观看| 边亲边吃奶的免费视频| 国产亚洲91精品色在线| 午夜爱爱视频在线播放| 欧美潮喷喷水| 在线天堂最新版资源| 亚洲欧洲国产日韩| 日日干狠狠操夜夜爽| 婷婷色av中文字幕| 亚洲无线观看免费| 日日啪夜夜撸| 我要看日韩黄色一级片| 国产中年淑女户外野战色| 亚洲丝袜综合中文字幕| 亚洲av日韩在线播放| 亚洲精品乱久久久久久| 亚洲综合精品二区| 欧美高清性xxxxhd video| 成人高潮视频无遮挡免费网站| 免费看a级黄色片| 视频中文字幕在线观看| 亚洲美女搞黄在线观看| 夫妻性生交免费视频一级片| 免费av毛片视频| 丰满少妇做爰视频| 十八禁国产超污无遮挡网站| 国产精品1区2区在线观看.| 成人国产麻豆网| 一边摸一边抽搐一进一小说| 晚上一个人看的免费电影| 国产白丝娇喘喷水9色精品| 亚洲自拍偷在线| 国产精品女同一区二区软件| 久久精品久久久久久久性| 黄色日韩在线| 菩萨蛮人人尽说江南好唐韦庄 | 深爱激情五月婷婷| 日本欧美国产在线视频| 又黄又爽又刺激的免费视频.| 久久精品91蜜桃| 一边亲一边摸免费视频| 热99re8久久精品国产| 日本熟妇午夜| 七月丁香在线播放| 18禁裸乳无遮挡免费网站照片| 乱码一卡2卡4卡精品| av.在线天堂| 久久久久久久久久黄片| 大香蕉97超碰在线| 免费观看精品视频网站| 欧美一区二区精品小视频在线| 国国产精品蜜臀av免费| 久99久视频精品免费| 亚洲精华国产精华液的使用体验| 午夜福利在线在线| 国产精品女同一区二区软件| 久久久国产成人精品二区| 网址你懂的国产日韩在线| 日韩精品有码人妻一区| 少妇猛男粗大的猛烈进出视频 | 日韩av在线免费看完整版不卡| 国国产精品蜜臀av免费| 久久人人爽人人片av| 亚洲,欧美,日韩| 亚洲怡红院男人天堂| 插阴视频在线观看视频| 久久精品熟女亚洲av麻豆精品 | 亚洲丝袜综合中文字幕| 久久99精品国语久久久| 亚洲av熟女| 国产av一区在线观看免费| 国产女主播在线喷水免费视频网站 | 少妇熟女aⅴ在线视频| eeuss影院久久| 禁无遮挡网站| 人人妻人人看人人澡| 国产欧美另类精品又又久久亚洲欧美| 成人无遮挡网站| 老师上课跳d突然被开到最大视频| 大话2 男鬼变身卡| 91久久精品国产一区二区三区| 久久久a久久爽久久v久久| 3wmmmm亚洲av在线观看| 国产精品美女特级片免费视频播放器| 中文字幕av在线有码专区| 国产真实伦视频高清在线观看| 毛片一级片免费看久久久久| 日韩一区二区视频免费看| 大话2 男鬼变身卡| 色吧在线观看| 国产精品熟女久久久久浪| 午夜免费男女啪啪视频观看| 免费播放大片免费观看视频在线观看 | 九九热线精品视视频播放| 婷婷色av中文字幕| 精品国产一区二区三区久久久樱花 | 人妻少妇偷人精品九色| 最近最新中文字幕大全电影3| 一个人看的www免费观看视频| av在线观看视频网站免费| 成人午夜高清在线视频| 国产 一区 欧美 日韩| 久久精品人妻少妇| 天美传媒精品一区二区| 别揉我奶头 嗯啊视频| 午夜福利在线观看免费完整高清在| 丰满乱子伦码专区| 如何舔出高潮| 夜夜看夜夜爽夜夜摸| 欧美一区二区国产精品久久精品| 美女高潮的动态| 小说图片视频综合网站| 男女边吃奶边做爰视频| 亚洲一区高清亚洲精品| 国产精品一区二区在线观看99 | 夫妻性生交免费视频一级片| 又粗又硬又长又爽又黄的视频| 热99re8久久精品国产| 国产精品国产高清国产av| 日韩欧美三级三区| 久久精品夜色国产| 在线观看66精品国产| 床上黄色一级片| 亚洲国产成人一精品久久久| 久久久久久九九精品二区国产| 国产黄色小视频在线观看| 一级爰片在线观看| 免费观看a级毛片全部| 久久久国产成人免费| 欧美性猛交黑人性爽| АⅤ资源中文在线天堂| 麻豆av噜噜一区二区三区| 亚洲av二区三区四区| 激情 狠狠 欧美| 99久久成人亚洲精品观看| 亚洲精品亚洲一区二区| 亚洲五月天丁香| 少妇的逼水好多| 精品一区二区免费观看| 非洲黑人性xxxx精品又粗又长| h日本视频在线播放| 97在线视频观看| 色尼玛亚洲综合影院| 美女大奶头视频| 精品无人区乱码1区二区| 内射极品少妇av片p| 又粗又硬又长又爽又黄的视频| 国产精品蜜桃在线观看| 日本-黄色视频高清免费观看| 校园人妻丝袜中文字幕| 99热精品在线国产| 日日撸夜夜添| 欧美高清性xxxxhd video| 国产精品福利在线免费观看| 国产精品熟女久久久久浪| 欧美激情久久久久久爽电影| 麻豆国产97在线/欧美| 亚洲久久久久久中文字幕| 久久久久久九九精品二区国产| 国产成人a∨麻豆精品| 波多野结衣巨乳人妻| 精品久久久久久电影网 | 天天一区二区日本电影三级| 亚洲18禁久久av| 晚上一个人看的免费电影| 日本免费一区二区三区高清不卡| 丝袜美腿在线中文| 麻豆一二三区av精品| 亚洲国产精品专区欧美| 日本黄色片子视频| 日本黄大片高清| 久久韩国三级中文字幕| 久久久欧美国产精品| 97超视频在线观看视频| 免费看日本二区| 免费观看在线日韩| 日日干狠狠操夜夜爽| 精品久久久久久久人妻蜜臀av| 色综合亚洲欧美另类图片| 成年免费大片在线观看| 亚洲第一区二区三区不卡| 丝袜美腿在线中文| 中文字幕av在线有码专区| 国产私拍福利视频在线观看| 久久精品国产99精品国产亚洲性色| 日本黄大片高清| 久久韩国三级中文字幕| 日韩成人伦理影院| 中文精品一卡2卡3卡4更新| 亚洲国产成人一精品久久久| 国产乱人视频| 亚洲国产最新在线播放| 搞女人的毛片| 国国产精品蜜臀av免费| videos熟女内射| av免费观看日本| 51国产日韩欧美| 亚洲国产高清在线一区二区三| 69人妻影院| 久久久欧美国产精品| 免费不卡的大黄色大毛片视频在线观看 | 久久99热这里只有精品18| a级毛色黄片| 边亲边吃奶的免费视频| 一边摸一边抽搐一进一小说| 天天一区二区日本电影三级| 成人特级av手机在线观看| 男女那种视频在线观看| 日韩精品有码人妻一区| 干丝袜人妻中文字幕| 男女啪啪激烈高潮av片| 春色校园在线视频观看| 女人久久www免费人成看片 | 天堂网av新在线| 少妇丰满av| 干丝袜人妻中文字幕| 日韩一本色道免费dvd| 2021少妇久久久久久久久久久| 网址你懂的国产日韩在线| 观看美女的网站| 18+在线观看网站| 亚洲内射少妇av| 亚洲人成网站高清观看| 九九久久精品国产亚洲av麻豆| 极品教师在线视频| 婷婷色麻豆天堂久久 | 亚洲欧美日韩东京热| 欧美性感艳星| 久久亚洲国产成人精品v| 有码 亚洲区| 久久久久久久亚洲中文字幕| 久久99热这里只有精品18| 中文字幕亚洲精品专区| 国产不卡一卡二| 一卡2卡三卡四卡精品乱码亚洲| 美女大奶头视频| 又爽又黄无遮挡网站| 中文欧美无线码| 久久久久性生活片| av在线观看视频网站免费| 非洲黑人性xxxx精品又粗又长| 中文字幕av成人在线电影| 久久久国产成人免费| 亚洲欧美精品自产自拍| 久久国内精品自在自线图片| 日本免费在线观看一区| 亚洲av免费高清在线观看| 身体一侧抽搐| 国产片特级美女逼逼视频| 日本-黄色视频高清免费观看| 国语对白做爰xxxⅹ性视频网站| 国产老妇女一区| 久久久久九九精品影院| 在线天堂最新版资源| 国产精品嫩草影院av在线观看| 天堂中文最新版在线下载 | 亚洲av成人精品一区久久| 我的女老师完整版在线观看| 国产大屁股一区二区在线视频| 亚洲av男天堂| 老女人水多毛片| 国产精品国产三级国产av玫瑰| 黄色欧美视频在线观看| 高清av免费在线| 看片在线看免费视频| 亚洲av福利一区| 免费观看精品视频网站| 99在线人妻在线中文字幕| 中文资源天堂在线| 亚洲国产精品成人综合色| 亚洲成色77777| 亚洲高清免费不卡视频| 国产午夜福利久久久久久| 中文字幕精品亚洲无线码一区| 欧美日韩精品成人综合77777| 九九爱精品视频在线观看| 九九久久精品国产亚洲av麻豆| av黄色大香蕉| 免费不卡的大黄色大毛片视频在线观看 | 嫩草影院精品99| 亚洲欧美精品专区久久| 国产精品国产三级国产专区5o | 国产精品熟女久久久久浪| 成人毛片a级毛片在线播放| 51国产日韩欧美| 亚洲欧美精品专区久久| 国产精品蜜桃在线观看| 亚洲欧美成人综合另类久久久 | 亚洲国产精品sss在线观看| 国产一级毛片在线| 日韩精品有码人妻一区| 老师上课跳d突然被开到最大视频| 七月丁香在线播放| 亚洲av成人精品一区久久| 亚洲内射少妇av|