• <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 ?

    VANET Jamming and Adversarial Attack Defense for Autonomous Vehicle Safety

    2022-08-24 03:30:28HaeriKimandJongMoonChung
    Computers Materials&Continua 2022年5期

    Haeri Kim and Jong-Moon Chung,2,*

    1Department of Vehicle Convergence Engineering,Yonsei University,Seoul,03722,Korea

    2School of Electrical&Electronic Engineering,Yonsei University,Seoul,03722,Korea

    Abstract:The development of Vehicular Ad-hoc Network(VANET)technology is helping Intelligent Transportation System (ITS) services to become a reality.Vehicles can use VANETs to communicate safety messages on the road(while driving)and can inform their location and share road condition information in real-time.However,intentional and unintentional(e.g.,packet/frame collision)wireless signal jamming can occur,which will degrade the quality of communication over the channel,preventing the reception of safety messages,and thereby posing a safety hazard to the vehicle’s passengers.In this paper,VANET jamming detection applying Support Vector Machine(SVM)machine learning technology is used to classify jamming and non-jamming situations.The analysis is based on two cases which include normal traffic and heavy traffic conditions,where the results show that the probability of packet dropping will increase when many vehicles are using the wireless channel simultaneously.When using SVM classification,the most appropriate feature set applied in determining a jamming situation shows an accuracy of 98%or higher.Furthermore,more advanced jamming attacks need to be considered for preparation of more reliable and safer autonomous ITS services.Such research can use vehicular communication transmission and reception data based on selected published datasets.In this paper,an additional adversarial defense algorithm using the Density-Based Spatial Clustering of Applications with Noise(DBSCAN)method is proposed,which assumes that evolutionary attacks of the jammer will attempt to confuse the trained classifier.The simulation results show that applying DBSCAN can improve the accuracy by elimination of outliers before conducting classification testing.

    Keywords: Vehicle safety;VANET;jamming;SVM;adversarial defense

    1 Introduction

    Recently, the development of self-driving and connected cars has led to the establishment of an Intelligent Transportation System (ITS) by connecting vehicles to the road infrastructure (V2I)(e.g., road-side units (RSUs) and cellular base stations), vehicles to vehicles (V2V), and vehicles to pedestrians(V2P),which form the overall description of vehicle to everything(V2X)networks[1].In addition,vehicular safety techniques are being developed to reduce the risk of road hazards,collisions,and accidents.The main goal of ITS is to make vehicles safer, faster, and more convenient to ride and drive.Vehicular ad-hoc network (VANET) technology is the basis for ITS.It was introduced in 2001 under vehicular ad-hoc mobile communication applications to provide road safety and roadside services[2].VANET applications include automatic toll systems,safety warning,platooning,and traffic information system support.In particular, vehicles performing time sensitive VANET applications for safety warning and driving guidance exchange urgent data to other vehicles and RSUs in a cooperative manner[3].This requires a VANET network with minimum data transmission delay and maximum reliability.

    In these VANET scenarios, intentional and unintentional (e.g., packet/frame collision) wireless signal jamming can occur.Intentional jamming attacks can seriously reduce the quality of the V2X communication performance[4].Jamming can cause the vehicle to suddenly stop or experience traffic jams due to traffic congestion.In addition,advanced jammers can attack as much as desired without detection as new undetectable attack patterns are developed.Therefore, much research on jamming detection has been conducted.

    However, mathematical estimation based on the correlation of representative communication parameters used in existing research has limitations.It is necessary to first identify the characteristics of various parameters in jamming situations and derive a parameter set that can obtain high detection accuracy.Therefore, in this paper, jamming detection is performed based on machine learning and the optimal communication parameter set is investigated.In addition,a detection algorithm against adversarial attacks is developed based on the assumption that the jammers continue to attempt evolved attacks.Before the detection test phase,a data pre-processing method is added to remove the data that makes classification difficult, based on the reason that the data points are located too close to the trained classifier or the noise which is far away.

    The sections of this paper are organized as follows.Details on VANET and jamming signals are provided in Section 2.Jamming detection using Support Vector Machine (SVM) technology is introduced in Section 3 along with corresponding experiment results.In Section 4, the proposed adversarial defense scheme that uses data sanitization to adapt to evolving attack patterns of the jammer is presented.Finally,the paper is concluded in Section 5.

    2 Background&Related Work

    2.1 VANET

    In support of V2I vehicle communications, the On-board Unit (OBU) and RSU communicate with each other,where the OBU is attached to the vehicle and the RSU is like a smart traffic light that can communicate with the vehicle.For road safety,all vehicles and base stations periodically send Basic Safety Messages(BSMs)at an average of 10 messages per second.BSMs include the vehicle’s location and incident information.Therefore,it is important to prevent the loss of the BSMs by reducing packet drops below the threshold level.

    2.2 Jamming Signal

    Jamming refers to a situation that an attacker intentionally sends a sequence of messages or signals to make vehicles unable to use the wireless channel [5].Or it makes the OBUs increase their transmission power so that other vehicles do not receive safety messages from other vehicles.The Signal to Noise Ratio(SNR)equationis computed by dividing the power of the signal by the noisen0and jamming signal powerJ0to determine the level of interference in the transmitted signal [6].Jamming has always been an important issue of investigation in communications, but especially,in vehicle communications it is more important as consistent connection to the RSU and other OBUs are critical to the safety of the driver and passengers.Because jammers can block safety messages,it can lead to a sudden stop,slow driving,traffic jams,and accidents.Therefore,if a jamming situation is detected, counter actions must be immediately taken, such as, reduce the impact on the jammer by increasing the power of the sending device to a certain level,or find the jammer’s location and/or attack pattern and make quick adjustments to avoid the jamming signal.

    2.3 Jamming Model

    Jammer attack types can be categorized according to mobility and behavior[4].Jamming mobility can be modeled as motionless(Stationary mobility),moving closer to a specific target vehicle(Target mobility), or moving around emitting a jamming signal without any target (Random mobility).In addition, the attack behavior can be divided into 3 patterns:Constant, Random, and Reactive.Constant behavior continuously transmits jamming signals, interfering with the channel access of regular vehicles by increasing the packet drop rate.The random pattern is when jammers randomly alternate between sleep mode and active mode to reduce their energy consumption and confuse the antijamming counter measure detection systems.Finally,a reactive pattern jammer senses a particular channel and attempts to attack only when a packet is transmitted.

    2.4 Related Work

    In recent research on VANETs, to improve the performance and provide adaptability, machine learning and deep learning optimization techniques have been applied to existing mathematical schemes.For example, for fast mobility and low energy consumption of vehicles, the authors of [7]develop a routing protocol based on reinforcement learning.Vehicles are clusted within the range of a RSU or a base station, and an optimal header node is selected for data collection and processing.Reinforcement learning adjusts the transmit power of each vehicle and continously motnitors the signal strength against the noise level.

    Related to the topic of this paper, the studies that focus on enhancing the jamming detection accuracy and avoidance techniques in VANETs are as follows.

    The authors of[8–10]propose a mathematical physical layer control scheme.The authors of[8]propose a new model based on the correlation among the error and the correct reception times to detect a jamming attack.The authors of [9] implement different jamming scenarios and evaluate jamming detection schemes,but only focus on RSU-OBUs based V2I communications.In addition,[10]focuses on the Bit Error Rate(BER)as the detection parameter.If the BER increases in time,the proposed scheme assumes that the nodes are moving towards a jamming area or adversary,then a plausibility check is done to check whether switching the frequency of communication is safe or not.

    The authors of [4] study jamming behaviors and effectiveness by focusing on 3 parameters,which are the Received Signal Strength(RSS),Packet Delivery Ratio(PDR),and Packet Send Ratio(PSR).The proposed scheme calculates the communication parameters based on the assumption that a jammer signal does not exist at initialization, and attempts to determine when jamming occurs.However, in real vehicular environments, every car generates lots of messages and such jammer-less initialization situations will not commonly occur in dense urban areas.

    Research in [11] implements a VANET intrusion detection application called the accurate and lightweight intrusion detection framework for vehicular networks(abbreviated as‘AECFV’).In this scheme, after clustering the vehicles, the individual detection algorithms for cluster heads, cluster members,and RSUs are conducted.In particular,at the cluster head level,anomaly detection and rulebased detection using SVM is performed considering 4 measurable parameters, which are the PSR,PDR,Message Duplication Ratio(MDR),and the Signal Strength Intensity(SSI).In use of the MDR,each cluster head is required to share its information with the RSU and needs other cluster members to perfrom their own periodic SVM training.Therefore, the RSU, cluster heads, and members in the VANET all need to have the AECFV application installed in order to conduct their required functions.As a result, the network overhead will significantly increase as the number of vehicles increases, caused by an increase in the number of messages they exchange in order to share control information.Unlike AECFV of[11],the scheme proposed in this paper only needs to be installed in the RSU.The parameters necessary for testing can be obtained periodically,and the presence of the jammer can be chekced at a RSU’s fixed position.

    The authors of [12] propose a jamming detection scheme that can be used to protect vehicle platoon applications.The purpose of jamming detection is to ensure that vehicles are not disturbed by a jammer by sending control messages to maintain the inter vehicle spacing.But, the detection algorithm only considers a constant jammer and the detection parameters used are the time delay between beacon messages and BER.

    In[13],a scheme that uses density-based spatial clustering of applications with noise(DBSCAN)and SVM to detect anomaly events using unsupervised learning in Wireless Sensor Network(WSN)is proposed.DBSCAN is used to label the sensor data based on clustering and then trains the system through SVM.Among the eight parameters used for training, the most important parameters for detection are temperature,humidity,and voltage.The results of[13]show that the propsoed scheme can effectively detect anomalies and recursively find the optimal DBSCAN parameter until the coefficient correlation(CC)value is minimized.

    Recently, the influence of jamming on Unmanned Aerial Vehicles (UAVs) networks has been studied in[14],where the authors propose a protocol-aware jammer avoidance scheme and implement a remote control system.Based on the communication indicator Jam to Signal Ratio(JSR),jamming signals with the characteristics of one,sweep,and protocol-aware are compared and analyzed based on the power required to successfully interfere with the network.

    The research of [15] proposes an UAV jamming detection scheme based on federated deep learning,where multiple clients and one central server collaborate,and a jamming detection module is integrated into the client UAV.In[15],the Received Signal Strength Indicator(RSSI)and PDR are used as training parameters to improve the learning performance,where a new client group prioritization method is introduced.

    In this paper,the jamming signal is detected using various communication parameters and a SVM machine learning detection system.The proposed scheme uses a sanitization algorithm that eliminates test points that are too far away(e.g.,noise factors)or test points that are too close to the SVM decision region,where its performance is compared to related antijamming schemes.

    3 Jamming Detection and Analysis

    In this section, a jamming detection method that uses SVM machine learning technology is proposed.Experiments are divided into two cases, where Case 1 has about 10 vehicles and 1 RSU,and Case 2 has more than 80 vehicles and 1 RSU.To increase the accuracy performance,the optimal feature sets and the kernel that can determine the jamming situation are searched.

    3.1 Support Vector Machine(SVM)

    SVM is a supervised machine learning algorithm which decides the boundary located midway between the training data classes [16].The linear discriminate is determined so that each class has a maximum margin.This margin is described as the vertical distance between the possible linear discriminator and the nearest points in either class.The data in either class which defines the maximum margin are referred to the support vectors that are presented in Fig.1,in whichzrepresents the training data points.The data points can be separated by a linear discriminatorH,andwis a normal vector perpendicular to the decision region, where the relation can be expressed aswz+b= 0, which is an equation of a straight line that isbaway from the origin.Therefore, based onH,H1is a minus plane,H2is a plus plane,and each of thezpoints can be distinguished by two classes(+1,-1).The dimension ofwdepends on the number of features used in the training,and if the number of features is more than 4,His represented by the word‘Hyperplane’because it is a decision region represented by a linear equation in a space of more than 4 dimensions.In this paper,SVM is used to classify normal and jamming signals.The research of [11,13] use SVM, which has an advantages of consuming less training time but provides a higher accuracy.

    Figure 1:Illustration of the SVM classification process

    In the case of this paper, an experiment was conducted based on (up to) 6 parameters (which is same as the 6 dimensions of the detection space) and SVM is used to separate the pattens in the hyperplane and make the detection process less sensitive to the number of points in each class[17].In addition,the kernel funcitonslinear,rbf,andpolynomialwere customized to enhance the accuracy of the detection.SVM was used in this paper because it is easier to control and tune compared to neural networks,and the classification performance is sufficiently good while the degree of over-fitting can be kept sufficiently low[18].

    3.2 Datasets

    The dataset used in this paper is the Dedicated Short Range Communication (DSRC) Vehicle Communications Dataset of [19] which is a freely distributed open-source dataset.This set includes data based on measurements of the wireless V2I and V2V communication.According to the Society of Automotive Engineers(SAE)Standard,DSRC sends 10 BSMs per second,which are 500 Byte User Datagram Protocol(UDP)packets[20].Measurement situations are based on a reactive jammer with 3 mixed mobility patterns(as described in Section 2.3)where the jammers only attack the control channel(Ch.127)which is one of the most important channels since safety messages are exchanged through it.Two separate datasets were used,which are the normal scenario(i.e.,non-jamming scenario)and the scenario that jamming attackers exist(i.e.,jamming scenario),thus each becomes a dataset which is labeled 0 and 1, respectively.In addition, a total of 1250 datasets are used and tested separately according to the vehicle traffic patterns.

    There are 6 main attributes of information on V2V car-to-car communication,which include the transmitted node’s ID number(Txnid),received node’s ID number(Rxnid),RSS(in dBm units),BER,RSSI,and SNR.The designed SVM scheme uses RSS and SNR as machine learning parameters in addition to PSR and PDR,which are monitored over the V2I RSU-to-vehicle communication.Further details of the parameters are provided below.

    3.3 Parameters

    Tab.1 describes the parameters to be used as the features of the SVM scheme.In studies [11]and[15],PDR and PSR values are calculated at each vehicle and there is a process of collecting data at the center vehicle and updating it.In this paper,each BSM includes the measured RSS and SNR values in every time message delivered,which is recorded in a log.In addition,information on,,,,andare collected through the packets received by the RSU every second.Finally,the PDR,PSR,and PDSR values are calculated using the equations of the parameters below.

    RSS:TheSmeasures the surrounding power of the receiver.If there is a jammer or traffic is congested because there are many vehicles,the signal strength could be high.LetPtbe the transmitter signal power,GtandGrare respectively the antenna gains of the transmitter and receiver,htandhrare respectively the height of the transmitter and receiver antennas,anddrepresents the distance between the transmitter and receiver.ThenScan be obtained from

    SNR:Theρrepresents the ratio of the signal power over the noise power,where jamming signals are a part of the noise.SNR is measured in units of decibels(dB).If the SNR is below the threshold value,the bit error probability increases,and the packet delivery ratio decreases,and the system fails to decode the signal.

    PDR:The Packet Delivery Ratio indicates the percentage of the correctly delivered packets,which is measured at the transmitter end.can be obtained fromandcan be obtained fromIf there is a jamming signal, the PDR level drops rapidly.During rush hours or jamming situations,packets will suffer from intentional interference causing a significant drop in the PDR.

    PSR:The Packet Send Ratio is the ratio of packets that are successfully sent out by a legitimate source compared to the number of packets that are intended to be sent out.can be obtained fromandcan be obtained fromWhen a jammer exists,the noise introduced by the jammer may hold the channel status busy,so more packets may be buffered and discarded upon the arrival of new packets.Therefore, using only the PSR as a feature alone can cause confusion about whether a channel is congested,or a jammer exists.

    PDSR:The Packet Delivery to Send Ratio is calculated by the RSU.PDSR is a parameter added to consider both the PSR and PDR,although CAR-PDSR and RSU-PDSR are not included in the existing dataset [6].can be obtained fromandcan be obtained fromIn an attempt to reduce the number of features of the SVM algorithm,the performance when using both the PSR and PDR will be compared to the performance when using the PDSR,as further described in the following.

    Table 1:Notations of parameters

    3.4 SVM Classification Method

    The proposed SVM scheme classifies jamming and non-jamming situations to determine the decision region and appropriate parameter/kernel set.For feature selection, the process of finding the highest jamming detection accuracy among the 4 feature sets are conducted.Sandρare used to identify the signal strength and channel noise state, thus the performance is compared to when using only PDR and PSR, and when replaced with PDSR.The experiments consider rush hour(vehicles ≥80)and normal traffic(vehicles ≈10)situations because having a dense scenario of vehicles can result in significant interference and performance degradation even without any jamming signal.The proposed SVM scheme uses the following basis procedures.

    1) Split the Dataset,Train:Test=8:2

    2) Feature Selection

    3) Kernel Selection

    a) Linear Kernel

    b) rbf Kernel

    c) Polynomial Kernel(degree 2 or 3)

    4) Train using the SVM model

    5) Classify the test dataset and analyze the results.

    3.5 Data Analysis

    To find the optimal feature set, an analysis on the accuracy of distinguishing the jamming situations based on the features was conducted.Fig.2 shows the correlation between the feature pairs and how the SVM decision boundary is drawn.First,RSS and SNR are significant factors that can easily determine the jamming situation as shown in Fig.2a.However,if there are many vehicles,the RSS increases and the SNR becomes small,and more similar to the jamming cases,so other features should be referred to.In the case of Figs.2b and 2c,PDR and PSR(which are measured by the RSU)are used as features,and the jamming situations are generally distinguished while allowing some errors.However,F(xiàn)igs.2d and 2e show that most of the PSR values are high even in jamming situations,so the start of message transmission is successful,but as the PDR level degrades,it is more difficult to deliver the message to the receiver.In addition,in both cases,the normal and jamming situations are not clearly dividable using a hard margin, so it is difficult to use this technique alone.As shown in Fig.2f, when only using the PDR value calculated from the RSU and vehicle, it is very difficult to distinguish the signal groups and make an accurate detection decision.

    Figure 2:Correlation between features.(a)RSS vs.SNR,(b)RSS vs.RSU-PDR,(c)RSS vs.RSU-PSR,(d)RSS vs.CAR-PDR,(e)RSS vs.CAR-PSR,and(f)RSU-PDR vs.CAR-PDR

    3.6 Experimental Results

    For feature selection,according to Fig.3,the results of applying different kernels are presented.

    Linear and polynomial kernels were considered in finding the most appropriate SVM kernels,where the rbf and second order polynomial kernels were able to achieve a high performance.The rbf kernel was applied to the training,and the performance was observed through iterative statements by substituting the values from 0.1 to 5 in order to find the optimal Gamma andCvalues.

    Figure 3:SVM kernel selection experiment results.(a) Linear kernel, (b) rbf kernel, (c) polynomial kernel(degree=2),and(d)polynomial kernel(degree=4)

    In addition, 4 different features sets were applied to find the appropriate parameter set as presented in Tab.2.For normal traffic,which represents relatively low density traffic environments,the parameter setresults in the highest accuracy.For high density traffic,the parameter setandresult in the same performance.Based on the tests conducted with the PDSR,results show that there are benefits in reducing the number of features from 6 to 4.So,the feature set(d)was applied.

    According to the feature set,the 4 dimensions classification results are represented by two 3 dimensions graphs.Fig.4a compares the terms ofand Fig.4b comparesin normal traffic situations.Likewise,F(xiàn)ig.5a compares the terms ofand Fig.5b compares the effect ofin heavy traffic jamming environments.

    For Case 1,as shown in Fig.4,the test accuracy was 99.2%and the false alarm factor was 1/65.When traffic is normal,jamming is easier to distinguish.For example,CAR-PDSR is clearly 10%or less and RSU-PDSR is lower than 5%when a jammer exists.For Case 2,as shown in Fig.5,the test accuracy was 95.3%and false alarm factor increased to 3/65.As the number of vehicles increases,the channel becomes worse and the SNR often falls below 40 dB, especially in the jamming situations.In general, jamming results in a higher RSS and lower SNR, because jamming signals act as noise.In the case of heavy traffic,the SNR difference between the jamming and normal scenarios was not significant.However, if the SNR is relatively high and the PDSR is low at the same time, jamming can be determined.The green‘x’points in Figs.4 and 5 are the false detection points,which are near the jamming decision region.As the green ‘x’points are very close to the classifier, they commonly make jamming detection fail.For Case 1, the average values of the green points result in a RSS of-68.823 dBm,SNR of 46.4398 dB,RSU-PDSR of 10.823%,and CAR-PDSR of 32.33%.On the other hand,the false alarm point of Case 2 results in a RSS of-58.8 dBm,SNR of 30.2 dB,RSU-PDSR of 54.5%,and CAR-PDSR of 46.53%.The most characteristic parameter that helps determine false alarm is RSS.Overall,the RSS ranges from-30 to-80 dBm,but when the received signal strength is not strong enough,it is close to the location of the classifier.Advanced adaptive jammers may try to attack as close as possible to the jamming decision region to prevent jamming detection.Therefore,in Section 4,an additional algorithm to increase the accuracy against evolved attacks is introduced.

    Table 2:Kernel accuracy

    Figure 4:Case 1 SVM test result.(a) RSS vs. RSU-PDSR vs. CAR-PDSR and (b) RSU-PDSR vs.CAR-PDSR vs.SNR

    Figure 5:Case 2 SVM test result.(a) RSS vs. RSU-PDSR vs. CAR-PDSR and (b) RSU-PDSR vs.CAR-PDSR vs.SNR

    4 Adversarial Attack and Defense

    In the previous section,SVM classification was applied to distinguish jamming and non-jamming situations.In many cases,when a jammer conducts an advanced attack,the VANET can use past attack information to detect the jammer’s behavior.For example, Wi-Fi based authentication procedures can be applied as public data records, and can be used for trainning based detection.In wireless communication environments,signal jamming attacks can occur anywhere and are very effective denial of service (DoS) attack tools, as the attacker is very difficult to find, attacks are much less costly,and these attacks have very low dependency to prior information compared to other spoofing and eavesdropping attacks[21].To defend against an advanced jammer,it is important to determine the jamming situation quickly and accurately before selecting an anti-jamming strategy in order to increase the success rate of the packet transmission.Therefore,an additional algorithm is proposed to prevent misclassification situations when an attacker evolves in the test phase.In order to defend against adversarial attacks,a data sanitization scheme using DBSCAN is proposed in this paper.

    4.1 Adversarial Attack Models

    Despite the high accuracy and performance of various machine learning algorithms, many schemes have been found to be vulnerable to subtle perturbation that has catastrophic consequences in security related environments.Adversarial Attacks can be categorized to several types[22].

    Poisoning Attackpatterns are known to contaminate the training data,which takes place during the training time.An adversary tries to poison the training data by injecting carefully designed samples to mislead the classifier.

    Evasion Attackoccurs by injecting malicious data or adversarial examples during the test phase.

    Blackbox Attackis used when the attackers do not know the inner configuration and attacks the models by continuously feeding samples and observing the output.

    Whitebox Attackoccurs when the adversary has access to all the information of the target neural network,including the model architecutre and weights,etc.

    In this paper, Evasion and Blackbox attacks are considered, where the attacker is constantly sending classifier confusing test data.

    4.2 Adversarial Sanitation Defense

    The attacker wants to cross the decision boundary towards the normal situation or place the points as close as possible to the boundary.So,the proposed adversarial defense is related to two ideas.First,as more attacks and defenses are repeated,the attacker will lower the PDR or/and PSR to the middle to look like a normal setup.Second, the jammer can’t go to the normal point perfectly because it’s presence alone acts as noise and lowers the SNR.In other words,the jammer can’t control the SNR parameter accurately.

    The proposed defense algorithm is based on data sanitization by using the DBSCAN algorithm[23].It can identify the sparse data as noise and form clustering groups.Assume the attacker usually has a fixed budget and if the dataset is large enough,adversarial examples would be sparse outliers.The algorithm operates to find the DBSCAN optimal parameters which are epsilons and MinPts.Epsilons are the size of neighborhoods and MinPts are the minimum number of points in the neighborhood.The scheme performs outlier removal on the test dataset using DBSCAN.To improve the performance of DBSCAN,a K-Means clustering technique could be used.However,K-Means requires the number of clusters to be set in advance[24],and the DBSCAN scheme determines the number of clusters to use during its classification process, without having this number preset in advance.Since DBSCAN uses the density,and the noise value does not belong to any cluster,it is easy to apply a pre-processing step.In addition,DBSCAN can also form clusters with an arbitrary shape.

    The proposed data sanitization scheme is analyzed using the vehicular parameters based on various characteristics,rather than trying to match the characteristics to a specific distribution shape[25].Research in[13]uses DBSCAN to label the train data,whereas this paper uses it to pre-process the test data and remove data points that can reduce the accracy of the trained classifier in advance.

    The scheme is tested separately when there are sufficient jamming data in the test samples and when there is little jamming data.Next,tests will be executed using the existing classifier on the cleaned data.

    Fig.6 illustrates the Data Sanitization process that pre-processes the data before the Test Phase.The condition of starting the initialization phase is to be able to adapt to changes in the vehicle density.Since the RSU collects road condition information(e.g.,number of vehicles on the road and location of the vehicles),it can renew its process whenever changes in vehicle density and jamming conditions are detected.For initialization, the initial test set is used to acquire MinPts and epsilon parameters required for DBSCAN.The initial test set is divided based on the vehicle density and already has labels.The optimal parameters values of MinPts and epsilon are the minimum values that satisfy the 90%test accuracy,which are obtained based on a recursively increasing algorithm of the MinPts and epsilon values.Gathering data from a test environment helps to eliminate outliers within the data that are suspected of jamming or confuse the classification.Therefore,a clean dataset is used for the test phase.

    Figure 6:Flow chart of data sanitization

    4.3 Results of Experiments

    The experiments that lead to the results presented in Figs.7 and 8 are based on the heavy traffic data,which has more than 80 vehicles with 1 RSU.According to the previously provided dataset,the false detection area among the existing jamming data patterns was randomly generated as a target.For example,RSS,SNR,CAR-PDSR,and RSU-PDSR sets for one point were moved close to the false detection region according to different random number values generated from a uniform distribution.Case 1 refers to a situation when there are sufficient jamming data.Jamming and normal test samples were respectively set to the same ratio of 50%and 50%based on a total of 630 samples.Fig.7 represents the DBSCAN results for Case 1,the orange dots refer to the non-jamming normal situation,the green dots represent the jamming situation,and the blue dots refer to jamming but difficult to classify region.The DBSCAN parameters epsilons and MinPts are 14 and 17,respectively.As shown in Fig.7,if the jamming test samples are sufficient,the attacker will insert a point mainly targeting the areas where errors occur,as it is close to the orange points,that is,the region which is difficult to determine.The jammer can adjust the number of jamming attacks and transmission power to make the RSS and PDSR parameters as similar as possible to the non-jamming situation,making detection difficult.As shown in the blue points region,the RSU-PDSR is similar to the non-jamming data,and the CARPDSR is close to jamming data at the same time.Since the RSU periodically collects road conditions and channel information to schedule the messages,RSU-PDSR tends to be more robust for jamming situations,as opposed to CAR-PDSR and RSU-PDSR,which can be measured accurately mostly in non-jamming situation.However, advanced attacks can also be detected by using the fact that the SNR is mostly low in the presence of a jamming signal.For example,the blue group has its RSS and PDSR located similar to the orange group,which reduces the detection accuracy to 96%.But clearing the blue group results in a detection test phase precision achieving 100%.

    Figure 7:Case 1 DBSCAN results

    Case 2 is a situation in which a small amount of jamming data is included in the test dataset.The total number of test samples is 420, of which contain less than 30% of jamming data samples.The DBSCAN parameters epsilons and MinPts for Case 2 are 14 and 8,respectively.The orange dots represent a non-jamming situation,and the green dots represent a jamming situation.If the jamming dataset is small, the number of points in a group will be small as well, where the points that are sufficiently far from the classifier can be removed (treated as outlier points), as shown in Fig.8.In the case 2 experiment, the blue dots refer to the mean noise.After removing theses data points, the scheme can be tested as a clean dataset.After the data sanitization process,the detection precision of the test phase resulted to be 99%.

    Figure 8:Case 2 DBSCAN results

    5 Conclusion

    In this paper,a jamming detection scheme to enhance the security of VANETs is proposed.When a jammer attacks,the measured values of the communication parameters,such as,RSS,SNR,PDR,and PSR change.Based on these parameters,SVM classification is performed using a dataset labeled with actual non-jamming and jamming data records.Furthermore, the results show that by using PDSR instead of using PDR and PSR helps to reduce the number of features, and a classification accuracy of 99%in normal traffic and 97%in heavy traffic can be obtianed.In addition,in this paper,a machine learning based adversarial defense algorithm for intelligent jamming detection is proposed.This scheme uses DBSCAN to detect intelligent jammer behavior that intentionally targets the decision region to reduce the accuracy of the SVM classification scheme.The DBSCAN based algorithm uses the fact that the SNR cannot be controlled by the jammer as the jamming signal is a type of noise.The clustering of DBSCAN allows the testing to be carried out with clean data, either by screening the jamming points near the decision region or by removing the outlier data points.

    The proposed jamming detection scheme is evaluated with 1 reactive jammer in the coverage of the RSU in this paper.To extend the evaluation of VANET jamming detection and scheme development,it would be benefcial to design the proposed scheme into an embedded network simulator module.Then,the specific jamming scenario with more reactive jammers can be tested in an automated jamming detection simulation environment.Therefore, it is desired for future work to extend the proposed scheme into an embedded vehicular network simulator and improve the scheme to perform as an enhanced automated detection module against more diverse attack scenarios.

    Funding Statement:This work was supported by the Institute for Information and communications Technology Promotion (IITP) grant funded by the South Korea government (MSIT, 2021-0-00040,Development of intelligent stealth technology for information and communication resources for public affairs and missions).

    Conflicts of Interest:The authors declare that they have no conflicts of interest to report regarding the present study.

    久久精品91无色码中文字幕| 中文欧美无线码| 大型黄色视频在线免费观看| 黑人欧美特级aaaaaa片| 中国美女看黄片| 18禁观看日本| 久久国产精品人妻蜜桃| 亚洲av国产av综合av卡| 欧美精品高潮呻吟av久久| 亚洲午夜理论影院| 国产成人影院久久av| 久久精品熟女亚洲av麻豆精品| 国产精品国产高清国产av | av有码第一页| 天天影视国产精品| 亚洲精品成人av观看孕妇| 国产精品偷伦视频观看了| 国产淫语在线视频| 亚洲精品成人av观看孕妇| √禁漫天堂资源中文www| 精品乱码久久久久久99久播| avwww免费| 交换朋友夫妻互换小说| 精品久久久久久电影网| 欧美在线一区亚洲| 欧美老熟妇乱子伦牲交| 久久久国产欧美日韩av| 日韩中文字幕欧美一区二区| 精品高清国产在线一区| aaaaa片日本免费| 国产黄频视频在线观看| 欧美日韩亚洲综合一区二区三区_| 亚洲,欧美精品.| 国产男靠女视频免费网站| 侵犯人妻中文字幕一二三四区| 精品高清国产在线一区| 久久国产精品人妻蜜桃| 婷婷丁香在线五月| 无人区码免费观看不卡 | 我的亚洲天堂| 成人三级做爰电影| 亚洲伊人久久精品综合| 高清毛片免费观看视频网站 | 国产免费福利视频在线观看| 美女高潮喷水抽搐中文字幕| 999久久久国产精品视频| 一级黄色大片毛片| 午夜福利视频精品| 精品一区二区三区av网在线观看 | 亚洲熟女精品中文字幕| 亚洲av第一区精品v没综合| 不卡av一区二区三区| 国产高清国产精品国产三级| 午夜福利在线观看吧| 人人妻人人爽人人添夜夜欢视频| 成人av一区二区三区在线看| 亚洲一区中文字幕在线| 国产成人av教育| 亚洲国产欧美一区二区综合| 十八禁网站网址无遮挡| 亚洲国产毛片av蜜桃av| 在线观看免费视频日本深夜| 久久久久国产一级毛片高清牌| 国产精品偷伦视频观看了| 国产成人精品久久二区二区91| 国产精品久久久人人做人人爽| 18禁观看日本| 一本综合久久免费| 最新美女视频免费是黄的| 中文字幕色久视频| 一级,二级,三级黄色视频| 丝袜美腿诱惑在线| 男女床上黄色一级片免费看| 国产在线观看jvid| 亚洲熟女毛片儿| 精品卡一卡二卡四卡免费| 高清在线国产一区| 一本色道久久久久久精品综合| 欧美日韩视频精品一区| 免费看十八禁软件| 国产伦人伦偷精品视频| 捣出白浆h1v1| 两个人免费观看高清视频| 欧美精品高潮呻吟av久久| 黑人巨大精品欧美一区二区蜜桃| 动漫黄色视频在线观看| 亚洲一区中文字幕在线| 9191精品国产免费久久| 久久99热这里只频精品6学生| 老司机午夜十八禁免费视频| 女警被强在线播放| 免费观看a级毛片全部| 免费看十八禁软件| 欧美精品一区二区大全| 国产成人精品久久二区二区91| 久9热在线精品视频| 欧美精品一区二区大全| 亚洲av日韩精品久久久久久密| av欧美777| 免费少妇av软件| 欧美亚洲日本最大视频资源| 中文字幕精品免费在线观看视频| 欧美大码av| 午夜91福利影院| 狠狠狠狠99中文字幕| 99在线人妻在线中文字幕 | 99国产精品99久久久久| 一进一出抽搐动态| 香蕉丝袜av| 亚洲视频免费观看视频| 午夜精品国产一区二区电影| 五月天丁香电影| 美女国产高潮福利片在线看| 搡老熟女国产l中国老女人| 午夜免费鲁丝| 黄色怎么调成土黄色| 国产一区二区 视频在线| 亚洲欧洲精品一区二区精品久久久| 亚洲精品国产一区二区精华液| av片东京热男人的天堂| 久久久国产成人免费| 99香蕉大伊视频| 麻豆国产av国片精品| 久久久久国内视频| 久久精品91无色码中文字幕| 中文字幕另类日韩欧美亚洲嫩草| 99国产精品免费福利视频| 热99久久久久精品小说推荐| 日本欧美视频一区| 久久久久久久国产电影| 热re99久久精品国产66热6| 亚洲va日本ⅴa欧美va伊人久久| 亚洲av成人一区二区三| 国产精品一区二区在线不卡| 精品一区二区三区四区五区乱码| 免费在线观看黄色视频的| 夜夜爽天天搞| 久久99一区二区三区| 亚洲人成77777在线视频| 国产xxxxx性猛交| av线在线观看网站| 飞空精品影院首页| 老汉色∧v一级毛片| 在线观看66精品国产| 女人爽到高潮嗷嗷叫在线视频| 91老司机精品| 黑人猛操日本美女一级片| 亚洲精品自拍成人| 最近最新中文字幕大全免费视频| 一夜夜www| av有码第一页| 两个人免费观看高清视频| 老熟妇仑乱视频hdxx| 精品国产亚洲在线| 免费在线观看影片大全网站| 在线观看www视频免费| 色婷婷久久久亚洲欧美| 国产精品熟女久久久久浪| 亚洲va日本ⅴa欧美va伊人久久| 色婷婷久久久亚洲欧美| 老司机在亚洲福利影院| 黄网站色视频无遮挡免费观看| 老司机在亚洲福利影院| 热99国产精品久久久久久7| 岛国在线观看网站| 国产精品一区二区精品视频观看| 老熟妇仑乱视频hdxx| 在线观看免费视频网站a站| 成年动漫av网址| 国产高清激情床上av| 亚洲,欧美精品.| 午夜福利,免费看| 午夜福利免费观看在线| 亚洲人成77777在线视频| 男女免费视频国产| 欧美精品一区二区大全| 大香蕉久久成人网| 国产精品秋霞免费鲁丝片| 久久精品国产99精品国产亚洲性色 | 丰满迷人的少妇在线观看| 国产成人啪精品午夜网站| 国产男女内射视频| 国产日韩一区二区三区精品不卡| 国产亚洲午夜精品一区二区久久| √禁漫天堂资源中文www| 最新美女视频免费是黄的| 50天的宝宝边吃奶边哭怎么回事| 免费一级毛片在线播放高清视频 | 国产不卡av网站在线观看| 热99久久久久精品小说推荐| 精品少妇黑人巨大在线播放| 少妇被粗大的猛进出69影院| 人妻 亚洲 视频| 美女午夜性视频免费| 操美女的视频在线观看| 久久精品国产亚洲av高清一级| 五月开心婷婷网| 日本a在线网址| 99精国产麻豆久久婷婷| 久久久欧美国产精品| 日韩欧美三级三区| 午夜福利,免费看| 久久国产精品男人的天堂亚洲| 亚洲精品在线美女| 大片免费播放器 马上看| 真人做人爱边吃奶动态| av国产精品久久久久影院| 人人妻人人添人人爽欧美一区卜| 成人三级做爰电影| 欧美另类亚洲清纯唯美| 青青草视频在线视频观看| 亚洲精品久久午夜乱码| 久久精品国产综合久久久| 黑人操中国人逼视频| 99国产精品99久久久久| 757午夜福利合集在线观看| 777久久人妻少妇嫩草av网站| 亚洲欧美日韩另类电影网站| 青青草视频在线视频观看| 欧美中文综合在线视频| 777米奇影视久久| 国产高清videossex| 欧美中文综合在线视频| 欧美久久黑人一区二区| 亚洲欧美一区二区三区久久| 97人妻天天添夜夜摸| 亚洲精品在线美女| 91成人精品电影| 国产av一区二区精品久久| 19禁男女啪啪无遮挡网站| 成在线人永久免费视频| 久久久久国内视频| 欧美日韩福利视频一区二区| 首页视频小说图片口味搜索| 搡老乐熟女国产| 9191精品国产免费久久| 又紧又爽又黄一区二区| 飞空精品影院首页| 美女午夜性视频免费| 欧美黄色片欧美黄色片| 性高湖久久久久久久久免费观看| 天天操日日干夜夜撸| 人人妻人人澡人人看| 2018国产大陆天天弄谢| 中文字幕av电影在线播放| 激情在线观看视频在线高清 | 在线永久观看黄色视频| 久久影院123| 亚洲自偷自拍图片 自拍| 十八禁高潮呻吟视频| 免费少妇av软件| 亚洲精品国产精品久久久不卡| 人成视频在线观看免费观看| 精品熟女少妇八av免费久了| 欧美中文综合在线视频| 91精品国产国语对白视频| 麻豆成人av在线观看| tube8黄色片| 精品少妇黑人巨大在线播放| 性色av乱码一区二区三区2| 99国产精品一区二区蜜桃av | 亚洲av第一区精品v没综合| 麻豆国产av国片精品| 午夜成年电影在线免费观看| 久久精品国产综合久久久| 午夜激情av网站| 午夜视频精品福利| 日韩欧美一区视频在线观看| 亚洲精品国产精品久久久不卡| 大型av网站在线播放| 国产一区二区在线观看av| 日本av手机在线免费观看| 国产1区2区3区精品| 女性被躁到高潮视频| 妹子高潮喷水视频| 国产真人三级小视频在线观看| 久久久久精品人妻al黑| 免费高清在线观看日韩| 欧美在线一区亚洲| 欧美日韩一级在线毛片| 亚洲成a人片在线一区二区| 国产高清videossex| 黄片播放在线免费| 国产99久久九九免费精品| 欧美激情久久久久久爽电影 | 久久午夜综合久久蜜桃| 欧美大码av| 热re99久久国产66热| 曰老女人黄片| 国产成人欧美| 丁香欧美五月| 一个人免费在线观看的高清视频| 超色免费av| 老司机午夜福利在线观看视频 | 欧美人与性动交α欧美软件| 亚洲三区欧美一区| 久久精品国产a三级三级三级| 亚洲成人国产一区在线观看| 叶爱在线成人免费视频播放| a级毛片在线看网站| 夜夜夜夜夜久久久久| 丰满人妻熟妇乱又伦精品不卡| 午夜日韩欧美国产| 成人av一区二区三区在线看| 精品一品国产午夜福利视频| 男女床上黄色一级片免费看| 男女免费视频国产| e午夜精品久久久久久久| 国产av国产精品国产| 一区二区三区激情视频| 美女午夜性视频免费| 90打野战视频偷拍视频| 久久午夜亚洲精品久久| 熟女少妇亚洲综合色aaa.| 黑人操中国人逼视频| 欧美 日韩 精品 国产| 国产xxxxx性猛交| 久久久水蜜桃国产精品网| 国产精品国产av在线观看| 久久久国产精品麻豆| 久久久国产成人免费| 久久人妻av系列| 九色亚洲精品在线播放| 国产日韩欧美在线精品| 久久中文字幕人妻熟女| 精品国产乱码久久久久久小说| 精品一区二区三区av网在线观看 | www.自偷自拍.com| 男女床上黄色一级片免费看| 欧美精品一区二区大全| 国产单亲对白刺激| 国产av国产精品国产| 国产亚洲一区二区精品| 91麻豆精品激情在线观看国产 | 热99国产精品久久久久久7| 日本五十路高清| 叶爱在线成人免费视频播放| 日韩一卡2卡3卡4卡2021年| 精品高清国产在线一区| 亚洲精品在线观看二区| 午夜福利在线免费观看网站| bbb黄色大片| 午夜福利视频精品| 乱人伦中国视频| 一个人免费在线观看的高清视频| 99精品久久久久人妻精品| 国产一区二区 视频在线| 午夜两性在线视频| 久久天堂一区二区三区四区| 亚洲精品自拍成人| 亚洲欧洲精品一区二区精品久久久| 成年人午夜在线观看视频| 国产又爽黄色视频| 在线看a的网站| aaaaa片日本免费| 一区二区三区乱码不卡18| 国产野战对白在线观看| 亚洲欧美一区二区三区黑人| 自拍欧美九色日韩亚洲蝌蚪91| 久久久久网色| 巨乳人妻的诱惑在线观看| 成人黄色视频免费在线看| 精品国内亚洲2022精品成人 | 18禁黄网站禁片午夜丰满| 精品国产国语对白av| 精品亚洲乱码少妇综合久久| 一本一本久久a久久精品综合妖精| 男人操女人黄网站| 国产99久久九九免费精品| 老熟妇仑乱视频hdxx| 91精品三级在线观看| 少妇的丰满在线观看| 五月开心婷婷网| 99香蕉大伊视频| 美女福利国产在线| 伊人久久大香线蕉亚洲五| 视频区欧美日本亚洲| 中文字幕人妻熟女乱码| 久久99热这里只频精品6学生| 久久久精品区二区三区| 亚洲熟妇熟女久久| 国产精品偷伦视频观看了| 99久久国产精品久久久| 黄色怎么调成土黄色| 丝袜在线中文字幕| 亚洲精品粉嫩美女一区| 一夜夜www| 菩萨蛮人人尽说江南好唐韦庄| 黑人欧美特级aaaaaa片| 在线观看舔阴道视频| 搡老岳熟女国产| 天天添夜夜摸| 99久久精品国产亚洲精品| 亚洲午夜理论影院| 波多野结衣av一区二区av| 狠狠狠狠99中文字幕| 法律面前人人平等表现在哪些方面| 下体分泌物呈黄色| 久久av网站| 18禁黄网站禁片午夜丰满| 一本大道久久a久久精品| 日韩 欧美 亚洲 中文字幕| 男女边摸边吃奶| 天堂8中文在线网| 丝瓜视频免费看黄片| 露出奶头的视频| 中文字幕色久视频| 亚洲一区中文字幕在线| 欧美黑人欧美精品刺激| 欧美日韩福利视频一区二区| av电影中文网址| 国产日韩欧美在线精品| 母亲3免费完整高清在线观看| 久久午夜综合久久蜜桃| 欧美日韩亚洲高清精品| 亚洲av日韩精品久久久久久密| 首页视频小说图片口味搜索| 成年版毛片免费区| 人妻一区二区av| 制服诱惑二区| 久久99热这里只频精品6学生| 成人国产一区最新在线观看| 国产亚洲精品久久久久5区| 夫妻午夜视频| 丝袜在线中文字幕| 亚洲人成77777在线视频| 老司机福利观看| 丰满饥渴人妻一区二区三| 天堂动漫精品| 亚洲人成伊人成综合网2020| 欧美精品啪啪一区二区三区| 1024香蕉在线观看| 国产欧美日韩综合在线一区二区| 天天躁日日躁夜夜躁夜夜| 一区福利在线观看| 久久九九热精品免费| 亚洲av成人一区二区三| 精品一品国产午夜福利视频| 久久性视频一级片| 丝袜美足系列| 国产主播在线观看一区二区| 亚洲色图 男人天堂 中文字幕| 精品国产国语对白av| 在线观看免费视频网站a站| 日本av免费视频播放| 午夜日韩欧美国产| av欧美777| 美女主播在线视频| av不卡在线播放| 蜜桃在线观看..| 国产精品秋霞免费鲁丝片| 美女国产高潮福利片在线看| 一本一本久久a久久精品综合妖精| 夜夜夜夜夜久久久久| 精品少妇一区二区三区视频日本电影| 久久久久精品人妻al黑| 桃花免费在线播放| 黄色视频在线播放观看不卡| 欧美日本中文国产一区发布| avwww免费| a级毛片在线看网站| av天堂久久9| 女人被躁到高潮嗷嗷叫费观| 日本a在线网址| 脱女人内裤的视频| 国产在线观看jvid| 后天国语完整版免费观看| 99国产精品一区二区三区| 亚洲成av片中文字幕在线观看| 国产一区二区在线观看av| 一边摸一边抽搐一进一出视频| 91精品三级在线观看| 一区在线观看完整版| 夫妻午夜视频| 老司机亚洲免费影院| 亚洲少妇的诱惑av| 91精品三级在线观看| 久久免费观看电影| 亚洲国产中文字幕在线视频| 免费在线观看视频国产中文字幕亚洲| 十八禁高潮呻吟视频| 国产精品一区二区在线观看99| 欧美日韩福利视频一区二区| 亚洲精品久久午夜乱码| 18禁观看日本| 精品免费久久久久久久清纯 | 国产日韩欧美在线精品| 精品国产一区二区三区四区第35| 男人舔女人的私密视频| 国产日韩欧美视频二区| 黄色成人免费大全| 一级黄色大片毛片| 午夜两性在线视频| 黑丝袜美女国产一区| 亚洲精品国产一区二区精华液| 丝袜在线中文字幕| 中亚洲国语对白在线视频| 免费一级毛片在线播放高清视频 | 在线观看人妻少妇| 欧美乱妇无乱码| 超碰成人久久| 成人免费观看视频高清| 久久影院123| 国产淫语在线视频| 色婷婷av一区二区三区视频| 成人av一区二区三区在线看| 1024视频免费在线观看| 中文字幕精品免费在线观看视频| 十八禁网站免费在线| 高清在线国产一区| 欧美av亚洲av综合av国产av| 视频区图区小说| 久久久久久免费高清国产稀缺| 纵有疾风起免费观看全集完整版| 精品卡一卡二卡四卡免费| 国产欧美日韩精品亚洲av| 色尼玛亚洲综合影院| 亚洲欧美色中文字幕在线| 日韩有码中文字幕| 国产亚洲一区二区精品| 高清黄色对白视频在线免费看| 亚洲av成人一区二区三| 国产欧美亚洲国产| 一级毛片电影观看| 国产又爽黄色视频| 性少妇av在线| 91字幕亚洲| 国产精品成人在线| 亚洲午夜精品一区,二区,三区| 91av网站免费观看| 一级毛片女人18水好多| 国产精品国产av在线观看| av网站在线播放免费| 中文字幕制服av| 国产色视频综合| 午夜激情av网站| 久久人妻熟女aⅴ| 9色porny在线观看| 国产精品久久久人人做人人爽| 老司机福利观看| 一个人免费看片子| 99热国产这里只有精品6| 精品一区二区三区视频在线观看免费 | 99久久人妻综合| 成年人黄色毛片网站| 国产精品国产高清国产av | 日本黄色日本黄色录像| 国产人伦9x9x在线观看| 一区二区日韩欧美中文字幕| 老司机福利观看| 国产成人系列免费观看| 欧美黄色淫秽网站| videos熟女内射| 亚洲伊人久久精品综合| av线在线观看网站| 午夜老司机福利片| 日韩制服丝袜自拍偷拍| 国产国语露脸激情在线看| 老汉色∧v一级毛片| 久久性视频一级片| 一二三四在线观看免费中文在| 久久久久久久精品吃奶| 岛国在线观看网站| 国产成人精品无人区| 国产精品久久久久成人av| 91av网站免费观看| 欧美 亚洲 国产 日韩一| 狠狠婷婷综合久久久久久88av| 又黄又粗又硬又大视频| 男女之事视频高清在线观看| 男男h啪啪无遮挡| 日韩三级视频一区二区三区| 午夜免费鲁丝| 久久久水蜜桃国产精品网| 国产99久久九九免费精品| 国产精品二区激情视频| 国产在线精品亚洲第一网站| 日本精品一区二区三区蜜桃| 黄色a级毛片大全视频| 亚洲av美国av| 日本av手机在线免费观看| 一二三四社区在线视频社区8| 脱女人内裤的视频| 国产一区二区激情短视频| 中文字幕高清在线视频| 久久精品人人爽人人爽视色| 色综合婷婷激情| 黑丝袜美女国产一区| 欧美日韩国产mv在线观看视频| 久久精品人人爽人人爽视色| 亚洲中文日韩欧美视频| 天天添夜夜摸| 午夜福利影视在线免费观看| 中文字幕另类日韩欧美亚洲嫩草| 国产在视频线精品| 久久精品国产亚洲av高清一级| 热99国产精品久久久久久7| 在线看a的网站| 久久久久网色| 麻豆av在线久日| 成人特级黄色片久久久久久久 | 亚洲综合色网址| 久久精品熟女亚洲av麻豆精品| 欧美国产精品一级二级三级| 久久狼人影院| 丰满人妻熟妇乱又伦精品不卡| 国产精品国产av在线观看| 久久ye,这里只有精品| 一本一本久久a久久精品综合妖精| 日本av手机在线免费观看| 亚洲av日韩精品久久久久久密| 欧美乱码精品一区二区三区| 国产亚洲精品久久久久5区| 亚洲av成人一区二区三| 少妇裸体淫交视频免费看高清 |