• <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 ?

    Reinforced CNN Forensic Discriminator to Detect Document Forgery by DCGAN

    2022-08-23 02:22:04SeoyoungLimandJeonghoCho
    Computers Materials&Continua 2022年6期

    Seo-young Lim and Jeongho Cho

    Department of Electrical Engineering,Soonchunhyang University,Asan,31538,Korea

    Abstract: Recently, the technology of digital image forgery based on a generative adversarial network (GAN) has considerably improved to the extent that it is difficult to distinguish it from the original image with the naked eye by compositing and editing a person’s face or a specific part with the original image.Thus,much attention has been paid to digital image forgery as a social issue.Further, document forgery through GANs can completely change the meaning and context in a document, and it is difficult to identify whether the document is forged or not,which is dangerous.Nonetheless,few studies have been conducted on document forgery and new forgery-related attacks have emerged daily.Therefore,in this study,we propose a novel convolutional neural network (CNN) forensic discriminator that can detect forged text or numeric images by GANs using CNNs, which have been widely used in image classification for many years.To strengthen the detection performance of the proposed CNN forensic discriminator, CNN was trained after image preprocessing,including salt and pepper as well as Gaussian noises.Moreover,we performed CNN optimization to make existing CNN more suitable for forged text or numeric image detection, which have mainly focused on the discrimination of forged faces to date.The test evaluation results using Hangul texts and numbers showed that the accuracy of forgery discrimination of the proposed method was significantly improved by 20%in Hangul texts and 5%in numbers compared with that of existing state-of-the-art methods, which proved the proposed model performance superiority and verified that it could be a useful tool in reducing crime potential.

    Keywords:Digital forensics;CNN;GAN;discriminator;image processing

    1 Introduction

    Owing to the COVID-19 pandemic, the use of non-face-to-face services through which official documents can be submitted and issued to public institutions has increased recently.According to a press release in 2020 by the Ministry of Security and Public Administration of the Republic of Korea,the number of annual usages of the“Mobile Document 24”application opened by the government has increased from 350,000 to 2,350,000 uses,more than 6.5 times since the first service launch in 2018.According to the increase in the usage rate of mobile official documents,mobile identifications have also been introduced,which are used to verify adult ages at convenience stores or for vehicle rentals through applications[1,2].

    However,the number of forged and falsified official documents continuously increases according to the data from Statistics Korea[2].Forged academic and grade transcripts of Bachelor’s,Master’s,and Doctorate degrees are used illegally for young people to get jobs.In illegal entry, residents’registration cards and passports are forged, as are endless crimes of official document forgery and falsification in society, eroding public trust and increasing social unrest [3].High-resolution digital cameras are now affordable,and software programs and source code are easily accessible,which further lowers the barriers to image manipulation.Thus, many techniques to detect fake images have been proposed to cope with easily generated fake images by adding,removing,or copying people and objects through relatively easy manipulation and editing.However,new forgery attacks are emerging daily[4].

    Recently,with the development of machine learning and artificial intelligence,images and videos can be manipulated digitally much easily than before.Moreover,an entirely new image can be created using its original image (Fig.1).Thus, the limitation of existing image manipulation can now be overcome.In particular, a generative adversarial network (GAN) is an unsupervised learning-based generative model implemented with two networks, which is regarded as one of the most promising technological developments in the image creation and manipulation field[5–8].Techniques to create data based on a GAN,such as image-to-image conversion[9–11]and Deepfake[12],which changes a person’s face into a different face to make a fake video,are frequently used to make fake images or videos.In addition, detailed manipulations can be possible by adjusting data in the image synthesis process,and a style may be applied arbitrarily[13,14].

    Figure 1: Image generated by stylegan2 ADA: The two images on the left are real images from the training set,and the last three images are generated through stylegan2 ADA

    Fig.2 shows an example of a fake document image for internal approval in the national administration office created using public administration documents of artificial intelligence hub.The left is the original image,and the right is the manipulated image.The total cost was artificially manipulated from KRW 1,774,000 to KRW 2,026,000.In this document,a fake numeric image was created through learning with deep convolutional GANs (DCGANs) using the same style of lettering, which was Gothic.This example demonstrates the possible risk of national budget waste through the forgery of unit prices or amounts in documents of national administration offices.Document forgery can completely change the context and meaning of the document in the real world and is difficult to be detected.Thus,a process to detect them efficiently is required.

    Figure 2: Document manipulated with numeric images created through a generative adversarial network(GAN);(a)original image(b)the image manipulated

    Recently,deep learning technology has been actively used in object recognition and classification fields,and a convolutional neural network(CNN)is particularly useful in finding a feature pattern for image recognition[15].As the trace of photo-response nonuniformity pattern remains in a shot image by a camera, various studies have been conducted to detect whether images generated with a GAN are forged using a CNN based on the fact that a specific fingerprint remains in an image created by a GAN[16].An 80%accuracy was realized in detecting forged images using the VGG16 architecture configured with five-layer blocks,including convolution and max-pooling layers[3],and the detection accuracy was improved to 91%using 13 convolution,5 pooling,and 3 dense layers by changing the architecture[17].However,since their proposed networks attempted learning using only forged images through a specific architecture, the detection performance of images forged by other architectures rapidly degraded [16].[18] employed a high-frequency pass filter (HFPF) to distinguish fake and real faces and determined whether the face is genuine through CNN using a residual.[19] detected fake images through CNN based on high-frequency components obtained through the HFPF after preprocessing via the Gaussian low-frequency pass filter, which suppresses the influence of external noise to identify face images generated by GAN that has undergone postprocessing such as image denoising and sharpening.In addition,the statistical similarity was improved at the pixel level between images through preprocessing using Gaussian blur and Gaussian noise to reinforce the generalization ability of the detection method of fake images created over other architectures,thereby learning more meaningful features inherently by the classifier[20].Consequently,the accuracy improved by up to 6%compared with that before using data preprocessing.However,the above previous CNN-based forgery discrimination techniques were designed for the digital synthesis image discrimination of human faces and did not verify the effectiveness of the discrimination of forged documents using a GAN,which is the aim of this study.Most recently,based on the fact that GAN-generated fake images have strong spectral correlations,[21]proposed a novel fake image detection strategy based on Discrete Wavelet Transform(DWT);GAN-generated images were transformed into three color components,R,G,and B, and then DWT and standard correlation coefficients were employed.As a result of the test, the detection accuracy was more than 90%, but the generalization performance was not considered as only styleGAN2-synthesized faces were included in the evaluation,and thus it was very limited.

    When a text in a forged document using a GAN is compared with that of the original document,it is not easy to distinguish it compared to when it is a face image.Thus,it is necessary to improve existing detection methods.Therefore,in this study,we propose a CNN-based detection strategy of document forgery created through learning with a GAN including preprocessing to improve the generalization ability of forgery discriminator and optimized CNN,by which existing CNN forensic discriminators apply to Hangul and numeric-based document forgery detection.The proposed strategy underwent preprocessing using salt and pepper noise pattern as well as Gaussian noise to improve the detection performance of forged documents.The performance evaluation results verified that the proposed method was effective in the forged text detection by training with a CNN using data containing various features.As a result of comparison with existing CNN-based fake face detection techniques,the proposed method was superior in forgery detection.This result verified the usability in effective discrimination of forged documents generated by a GAN,which are receiving significant attention in the field of digital image forensics recently.Further,the proposed method will significantly contribute to not only document forgery detection but also Deepfake detection research as a basic technology in the future.

    Fake news that provides false information through forged images, as well as information such as highly realistic fake profiles in a social network through the forgery and falsification of texts or numbers,can be harmfully used in other applications,which must be solved.Thus,the contributions of this study in this context are as follows:

    i.An optimization method of the CNN forensic model is proposed to extend existing discrimination techniques,which were limited to face images generated by a GAN,to the discrimination of Hangul and number forgery detection.

    ii.A data preprocessing technique is proposed to improve the diversity of GAN-generated data and the generalization ability of CNN forensic models.

    The rest of this paper is organized as follows:In Section 2,the data creation and proposed forgery document detection technique are described;in Section 3,experimental results are presented.Finally,in Section 4,conclusions are presented.

    2 Methodology

    Owing to the need for developing a detection technology to match with the growth rate of forgery technology, we propose an algorithm to detect texts, particularly Hangul or numeric forged images generated by a GAN.A CNN, which has been used to detect forged images, extracts input image features, and classifies images using the extracted pattern.However, because the generated forged image features may differ depending on the GAN’s architecture,sufficient images generated by various GAN models should be included in CNN learning to ensure the generalizability of CNN classifiers,but it is difficult to implement this.To solve this problem,we improved the generalizability of detectors by removing fingerprints of GANs through preprocessing of training images.Further,existing CNN detectors were mainly used as tools for fake face detection.Thus, we optimized CNN classifiers to realize forged texts or numbers using a GAN.The proposed reinforced discrimination strategy of GAN-generated forged document is shown in Fig.3, which comprises three sections: 1) fake image generator,2)preprocessor,and 3)CNN forensic discriminator.

    Figure 3:Overall framework of the proposed reinforced CNN forensic discriminator

    2.1 Fake Image Generator

    Because fake images made by a fake image generator are produced by their original images,original images were first created with a size of 28×28 pixels, which were similar fonts supported by the Windows.Here, the numbers were created with a size of 0.7 points and “2”bold using five fonts available in Python:“Italic,”“Hershey complex,”“Hershey duplex,”“Hershey script complex,”and “Hershey.” For Hangul, five fonts, “Gulim,” “Gungseo,” “Malgun Gothic,” “Batang,” and“Headline,”were used, and 14 consonant letters,were created with a size of 20 points.Fig.4 shows examples of original Hangul and numeric images in various fonts.

    Figure 4:Examples of original Hangul and numeric images in various fonts

    The DCGAN was used to create forged images for CNN learning and test evaluations based on the produced original images.The DCGAN is a GAN comprising two networks with a CNN structure.The first network is a generator,which creates fake images with the input of random noise,and the second network is a discriminator, which determines whether the generated image is real or fake.The generated fake image is input to the discriminator along with the original image to determine whether it is genuine or not.Based on this result,the generator is trained adversarially to generate more realistic fake images.Here,the training continues until it can fool the discriminator completely.The loss function is presented in Eq.(1), which aims for complete mapping of latent spaceZinto data space X.

    whereG(z)is the output of the generator, given a random noisez,D(x)is the estimate from the discriminator, Pdata(x)is the distribution of the original image, and Pz(z)refers to the Gaussian distribution of the latent space.The generator is trained to make the objective function V(D,G)the minimum,i.e.,D(G(z))= 1,whereas the discriminator is trained to make the objective function the maximum,i.e.,D(x)=1, D(G(z))=0.

    2.2 Preprocessor

    A fake image generated by the DCGAN provides a realistic image similar to the original image,but it can be easily distinguished from the original image because it has a specific fingerprint,thereby obtaining high accuracy.However,for images created by GAN models other than the DCGAN,which have the same architecture, the performance of discriminating whether a fake image is genuine is significantly degraded.To solve this drawback,salt and pepper as well as Gaussian noises were added to be used in the learning of discriminator to prevent the detection model from being more focused on training artificial fingerprints inserted by the GAN and image background than on a text itself.

    Salt and pepper noise randomly alters a certain amount of pixel into two incorrect brightness values,either 0 or 255,which occurs as a form of black or white point in an image.Generally,in the document background, visible watermark, anticounterfeiting pattern, and background noise can be included, which become the background behind a series of texts in a document.Thus, by adding a random salt and pepper noise pattern to the original image, the background is made in the form of noise instead of white,thereby allowing the CNN forensic discriminator to learn by focusing more on the text itself rather than the background.The amounts of salt and pepper were added equally at 50%each, and the added amount was randomly set between 0 and 0.5.Moreover, the statistics of image pixels were modified by adding the Gaussian noise of zero mean unit variance to make it difficult for DCGAN to focus only on detection through artificial fingerprints in training the CNN forensic discriminator.Preprocessed data, which are randomly sampled, are shown in Fig.5; we determined whether they are genuine as they enter the forensic detection unit.

    Figure 5:Examples of Hangul and numeric images with noise added through the preprocessor

    2.3 CNN Forensic Discriminator of Document Forgery

    Deep learning-based technologies have been widely used in object and face recognition as well as image classification recently.Among them,a CNN is particularly outstanding in finding a pattern to recognize images.A filter is applied to the input image,and a feature map,which is the same as that of the input image,is generated through convolution and combination,thereby classifying images using the extracted feature patterns.Owing to the advantage of automatically extracting image features,CNNs have shown the effectiveness to identify forged images [22].Thus, a CNN was considered the discriminator architecture to determine whether a document was forged,and the CNN detection model for discriminating forged text or number by GANs was optimized with the structure presented in Tab.1.

    Table 1: Proposed CNN architecture for discriminating document forgery

    For the network optimization, real and fake images were randomly mixed at a 1:1 ratio and extracted to have 100,000 and 20,000 images for learning and testing, respectively.To shorten the learning time and reduce the possibility of falling into the local minimum,all data were normalized,and the network had three convolution layers.All convolution layers employed 2×2 max pooling and“batch normalization.”The size of all convolution kernels was 3×3;the activation function used in each convolution layer was “ReLU”, but the last layer employed “SoftMax.”The loss function and optimization algorithm were sparse categorical cross-entropy loss function and adaptive momentum estimation(Adam).The loss function is shown in Eq.(2):

    where βiis the actual value,andαiis the estimated one.Adam was adopted to the optimization because it has few memory requirements as it can perform efficient operations through simple implementation.As the learning progresses, the learning rate was reduced, and the speed was calculated to adjust the learning intensity.In the forensic CNN model’s learning process,“Dropout”was used to prevent overfitting of training data.

    3 Experimental Results

    The forgery performance of the GAN was first verified before the performance evaluation of the proposed GAN-generated forgery document discriminator was performed.The hardware used for learning GAN and others included an Intel Core i7-7700 CPU, NVIDIA GTX 1060 GPU (6 GB),and 16 GB of memory.The software environment comprised Tensorflow 2.2.0,Opencv 4.5.1,CUDA v.10.2,and Cudnn v.7.6.4 on Ubuntu 16.04.5(4.15.0–38 kernel).The GAN-generated forged images were highly similar to the original images,which could not be clearly distinguishable by naked eyes.Thus,the following Fréchet inception distance(FID),which is a method of measuring the similarity between curves by considering the location and order of points along the curve,was used to calculate a distance of feature vectors between the original and fake images for the quality evaluation of how the forged image was close to the original image.

    where(μr,Covr)and(μg,Covg)are the mean and covariance of each original and generated image feature,respectively,and Tr is the trace operation.The lower the FID is,the better the forged image quality is, and the generated image is closer to the original image.The performance of the forged image generated using the DCGAN was compared with that of the Wasserstein GAN(WGAN),which is designed to compensate for the weakness of being difficult to train while maintaining a balance between the discriminator and generator of GAN.WGAN is a structure that uses the critic to deliver the gradient well and trains the critic and generator to the optimal point so that it can play a better role in classification than the discriminator of GAN.The comparative results are summarized in Tab.2.

    Table 2: FID to evaluate GAN-generated fake image performance

    Both techniques verified that as the FID was closer to zero, they could fake images better, and the FID of WGAN was lower in both number and Hangul than that of DCGAN,indicating that the images created by WGAN were closer to the original images.That is,it can be expected that the image generated by DCGAN will be relatively easier to detect forgery than the image generated by WGAN.Thus, DCGAN-generated images were used for CNN training to have more objective evaluations,whereas,for performance evaluation,WGAN-generated images were used to make closer images to the original images for harder discrimination.A comparison of performance evaluation on forged image detection according to various preprocessing procedures was performed.Tab.3 summarizes the models used in the comparison evaluation.In this study, based on an existing CNN forensic model optimized to face images, Mface, and a CNN model optimized to Hangul and numeric images, M, a CNN model trained through preprocessing by adding Gaussian noise to model M, a CNN model trained via preprocessing by adding the Gaussian blur noise to model M,and a CNN model trained additionally through preprocessing using the Gaussian noise as well as salt and pepper noise were defined as Mg,Mgb,and Mg+sp(proposed model),respectively.

    Table 3: Definition of models used in test evaluation

    For indices to evaluate the performance of the proposed CNN forensic model, accuracy, which calculates a probability of being classified correctly considering both the case of correctly predicting the real image as“True”and the case of correctly predicting the fake image as“False,”true negative rate(TNR),which is a rate of forged images predicted as fakes,and true positive rate(TPR),which is a rate of predicting the original image as the original.They can be calculated as follows:

    where TP means true positive,which is correctly determined case of original images,TN means true negative,which is correctly determined case of forged images.FP means false positive,which is a case where original images are wrongly determined as forged images,and FN means false negative,which is a case where fake images are wrongly determined as original images.

    Tab.4 presents the comparison results of detection performance among forensic models with and without preprocessing.Data used in the performance evaluation were numbers and Hangul,and forensic models M and Mg+spwere optimized through data comprising fake and original images generated by DCGAN.To verify the generalizability,forged data created using DCGAN,which was the same form used in training the CNN forensic model,and WGAN,which was a new form without participation in training,were employed.The performance results by forensic models with and without preprocessing using test data created by DCGAN showed that M and Mg+sphad no significant difference in forgery discrimination at a nearly similar level.However,the accuracy results using test data created by WGAN revealed that for numbers,the accuracies of M and Mg+spwere 76.44%and 82.07%,respectively,which showed the proposed model was 5.63% superior, and for Hangul, their respective accuracies were 84.94% and 88.22%, which showed the proposed model was approximately 3% superior.Thus, the above results verified that the proposed forensic model improved the generalizability and detection performance of fake images by incorporating preprocessing, focusing more on characters’feature analysis rather than on fingerprints of GAN or background of Hangul and numeric characters.

    Table 4:Performance comparisons of forensic models optimized for Hangul and numbers with and without preprocessing

    The comparative evaluation was performed to verify how much the proposed forgery discrimination structure was better in character detection than that of other state-of-the-art models.Tab.5 summarizes the results.Mfaceis an optimized forensic model for face detection,which is widely used in Deepfake detection, and Mgand Mgbare preprocessed forgery and face detection models using Gaussian noise and Gaussian blur noise.For numbers, the accuracy of Mfaceusing the DCGANand WGAN-generated test data was higher than that of Mgband Mgbut lower than that of Mg+spby 5.55% and 6.08%, respectively.In terms of execution flows, Mgband Mgwere similar to Mg+spin that they performed image preprocessing followed by CNN training, but they were focused on face identification.Thus, their accuracies degraded by 23.46% and 27.99%, compared with that of Mg+spwhen testing the generalizability usings WGAN-generated images, indicating that they were highly vulnerable to number detection.Moreover, for Hangul, the accuracy of Mg+spin the generalizability test using WGAN-generated images showed 36.59%,21.8%,and 21.95%better than those of Mface,Mg,and Mgb,respectively,which verified that the proposed method was much more suitable to Hangul and number detection forged with a GAN,showing robustness to new forms,which were GAN-generated images.

    Table 5:Performance comparisons of the proposed discriminator with state-of-the-art CNN forensic models

    Table 5:Continued

    Tabs.6 and 7 present examples of detection results by forensic models for numbers and Hangul,respectively.The first row shows the test data images,and the second row indicates the image is genuine or fake.The original image that is not faked is indicated as T(True)and the forged image as F(False).T or F cannot be distinguished by naked eyes,but the proposed forensic model can discriminate T or F of Hangul and numbers correctly.

    Table 6: Examples of detection results by forensic models for numbers

    Table 7: Examples of detection results by forensic models for Hangul

    4 Conclusions

    Recently,considerable attention has been paid to forgery and falsification using a GAN,which is rapidly advancing,and various techniques to detect them have been proposed.However,most studies were focused on face detection only.The forgery of documents using a GAN can completely change the context and meaning of the document and is much easier than that of a fake face,which makes it more difficult to distinguish from the original images.As a result,the crime rate of document forgery is steadily increasing.Thus,a new method to detect document forgery has been constantly demanded.In this context, this study performed an optimization process of the CNN forensic model to detect characters(Hangul and numbers)and increased the reliability of the possibility to detect text forgery.To overcome the shortcoming that the generalizability of models degrades due to specific fingerprints in images when detecting GAN-generated images using a CNN, changes in pixel values were given using the Gaussian Noise Pattern,and an image preprocessing process was proposed by adding a salt and pepper noise pattern,which could play a similar role as visible watermarking or anticounterfeit noise of documents.To verify how much more the proposed forgery discrimination structure was effective in character detection than that of other state-of-the-art models, comparative evaluations were performed on the basis of WGAN-generated numbers and Hangul test data,which have never been used in learning before.The evaluation results showed that the accuracy improved by more than 5%and 20%for numbers and Hangul,respectively,which verified that the proposed forgery detection structure was suitable for detecting Hangul and numbers forged by a GAN, and also demonstrated high robustness against new types of GAN-generated images.

    Acknowledgement:The authors thank the editor and anonymous reviewers for their helpful comments and valuable suggestions.

    Funding Statement:This research was funded by a National Research Foundation of Korea (NRF)grant funded by the Korean government(MOE)(No.2021R1I1A3055973)and the Soonchunhyang University Research Fund.

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

    久久久国产成人免费| 久久精品夜色国产| 久久国产乱子免费精品| 国产精品国产高清国产av| 看免费成人av毛片| 国产精品乱码一区二三区的特点| 亚洲在线自拍视频| 国内精品久久久久精免费| 国产美女午夜福利| 男插女下体视频免费在线播放| 少妇丰满av| 国产精品一区二区三区四区免费观看 | 亚洲最大成人中文| 日本黄色片子视频| 成人无遮挡网站| 午夜亚洲福利在线播放| 国产亚洲91精品色在线| 久久精品国产自在天天线| 国产三级中文精品| 国内精品美女久久久久久| 亚洲性久久影院| 亚洲不卡免费看| 午夜激情欧美在线| 性色avwww在线观看| 一级毛片电影观看 | 国产精品久久久久久精品电影| 男人舔奶头视频| 国产精品亚洲一级av第二区| 尾随美女入室| 国产在视频线在精品| 一级av片app| 乱系列少妇在线播放| 国产精品久久久久久亚洲av鲁大| 看片在线看免费视频| 国产精品一二三区在线看| 一本一本综合久久| 久久韩国三级中文字幕| 97超视频在线观看视频| 日日啪夜夜撸| 成人毛片a级毛片在线播放| 欧美不卡视频在线免费观看| 一个人看视频在线观看www免费| 国产精品野战在线观看| 欧美一级a爱片免费观看看| 国产又黄又爽又无遮挡在线| 国产大屁股一区二区在线视频| 午夜激情福利司机影院| 97热精品久久久久久| 噜噜噜噜噜久久久久久91| av视频在线观看入口| 国产精品久久久久久av不卡| 久久久精品94久久精品| 日韩强制内射视频| 久久精品国产亚洲av香蕉五月| 欧美日本亚洲视频在线播放| 天天躁夜夜躁狠狠久久av| 寂寞人妻少妇视频99o| 婷婷精品国产亚洲av| 女同久久另类99精品国产91| 日韩,欧美,国产一区二区三区 | 久久人人精品亚洲av| ponron亚洲| 国产精品久久久久久久电影| 成年版毛片免费区| 国产精品国产高清国产av| av在线亚洲专区| 国产精品人妻久久久久久| 精品久久久久久久久久免费视频| 亚洲精品乱码久久久v下载方式| 一级av片app| 亚洲一区高清亚洲精品| 久久精品国产亚洲网站| 三级毛片av免费| 精品人妻熟女av久视频| 国产真实乱freesex| 老司机影院成人| 欧美激情在线99| 日日干狠狠操夜夜爽| 亚洲性夜色夜夜综合| 女生性感内裤真人,穿戴方法视频| 免费av毛片视频| 日韩制服骚丝袜av| 精品乱码久久久久久99久播| 深夜精品福利| 波多野结衣高清作品| 少妇的逼好多水| 国产精品日韩av在线免费观看| 日韩强制内射视频| 中国国产av一级| 超碰av人人做人人爽久久| 狠狠狠狠99中文字幕| 欧美日韩乱码在线| 综合色av麻豆| 亚洲无线在线观看| 麻豆国产97在线/欧美| 精品久久国产蜜桃| 国产一区亚洲一区在线观看| 一进一出好大好爽视频| avwww免费| 成人三级黄色视频| eeuss影院久久| 日韩欧美三级三区| 久久99热这里只有精品18| 午夜a级毛片| 午夜影院日韩av| 五月伊人婷婷丁香| 日韩三级伦理在线观看| 国产高清视频在线观看网站| 亚洲自偷自拍三级| 国产男人的电影天堂91| 美女xxoo啪啪120秒动态图| 久久午夜福利片| 秋霞在线观看毛片| 国产日本99.免费观看| 久久久久性生活片| 蜜桃亚洲精品一区二区三区| 最新在线观看一区二区三区| a级毛色黄片| 欧美高清性xxxxhd video| 高清毛片免费看| 国产精品人妻久久久久久| 女同久久另类99精品国产91| 亚洲精品日韩在线中文字幕 | 久久欧美精品欧美久久欧美| 国产精品福利在线免费观看| 日本色播在线视频| 搡女人真爽免费视频火全软件 | 日韩成人av中文字幕在线观看 | 国产亚洲精品综合一区在线观看| 亚洲欧美日韩东京热| 亚洲经典国产精华液单| 免费观看在线日韩| 男女啪啪激烈高潮av片| 日韩精品青青久久久久久| 此物有八面人人有两片| 久久欧美精品欧美久久欧美| 无遮挡黄片免费观看| 五月伊人婷婷丁香| 热99在线观看视频| 精品午夜福利在线看| 国产成年人精品一区二区| 国产一区亚洲一区在线观看| 亚洲精品影视一区二区三区av| 亚洲高清免费不卡视频| 亚洲国产精品成人久久小说 | 波多野结衣高清作品| 国产黄a三级三级三级人| 亚洲精品粉嫩美女一区| 最近中文字幕高清免费大全6| 一个人观看的视频www高清免费观看| 69av精品久久久久久| 国产精品久久视频播放| 99在线视频只有这里精品首页| 成人av在线播放网站| 97超级碰碰碰精品色视频在线观看| 国产v大片淫在线免费观看| 亚洲精品456在线播放app| 欧美极品一区二区三区四区| 久久久精品大字幕| 搞女人的毛片| 久久久久国产网址| 网址你懂的国产日韩在线| 不卡一级毛片| 18禁黄网站禁片免费观看直播| 国产伦精品一区二区三区视频9| 久久久久久久久大av| a级毛片a级免费在线| 亚洲av一区综合| 欧美在线一区亚洲| 一个人观看的视频www高清免费观看| 久久久久国内视频| av女优亚洲男人天堂| 老司机福利观看| 日韩在线高清观看一区二区三区| 国产精品1区2区在线观看.| 大型黄色视频在线免费观看| 最好的美女福利视频网| av国产免费在线观看| 亚洲无线在线观看| av卡一久久| 99热这里只有是精品在线观看| 久久久久久久久久久丰满| 日日摸夜夜添夜夜添小说| 亚洲色图av天堂| 亚洲第一区二区三区不卡| 男人舔奶头视频| 久久这里只有精品中国| 国产精品日韩av在线免费观看| 一级毛片我不卡| 最近手机中文字幕大全| 免费观看的影片在线观看| 亚洲国产精品成人久久小说 | 久久6这里有精品| 麻豆国产97在线/欧美| a级毛片免费高清观看在线播放| 国产精品女同一区二区软件| 亚洲欧美日韩卡通动漫| 日日干狠狠操夜夜爽| 一区二区三区免费毛片| 好男人在线观看高清免费视频| 久久久国产成人免费| 国产中年淑女户外野战色| 成熟少妇高潮喷水视频| 精品久久久久久久久久免费视频| 熟妇人妻久久中文字幕3abv| 亚洲一区二区三区色噜噜| 精品一区二区免费观看| 成人漫画全彩无遮挡| 亚洲精品影视一区二区三区av| 国产精品伦人一区二区| 亚洲三级黄色毛片| 91在线观看av| 欧美日韩综合久久久久久| 精品久久久久久久久亚洲| 综合色av麻豆| 性色avwww在线观看| 午夜激情欧美在线| 老师上课跳d突然被开到最大视频| 亚洲国产精品合色在线| 免费在线观看影片大全网站| 国产高清激情床上av| 亚洲最大成人av| 亚洲va在线va天堂va国产| 如何舔出高潮| 免费人成在线观看视频色| .国产精品久久| 欧美日本亚洲视频在线播放| АⅤ资源中文在线天堂| 日本a在线网址| 别揉我奶头~嗯~啊~动态视频| 亚洲人成网站在线播| 免费人成在线观看视频色| 一级黄色大片毛片| 成人国产麻豆网| 日韩成人av中文字幕在线观看 | 免费av观看视频| 国产高清视频在线播放一区| 国产亚洲91精品色在线| 国内精品宾馆在线| 99久国产av精品| 国产精品精品国产色婷婷| 成人av在线播放网站| 99热精品在线国产| 亚洲欧美精品自产自拍| 欧美bdsm另类| 我要搜黄色片| 久久精品夜色国产| 国产精品野战在线观看| 永久网站在线| 俄罗斯特黄特色一大片| 最后的刺客免费高清国语| 色av中文字幕| 国产免费男女视频| 你懂的网址亚洲精品在线观看 | 免费在线观看成人毛片| 中文字幕免费在线视频6| 免费人成在线观看视频色| 国产白丝娇喘喷水9色精品| 国产精品福利在线免费观看| 亚洲性夜色夜夜综合| 国产麻豆成人av免费视频| av黄色大香蕉| 国产探花极品一区二区| av在线观看视频网站免费| 免费电影在线观看免费观看| 中国美女看黄片| 亚洲精品456在线播放app| 97超碰精品成人国产| 国产成人精品久久久久久| 搡女人真爽免费视频火全软件 | 卡戴珊不雅视频在线播放| 桃色一区二区三区在线观看| 在线国产一区二区在线| 亚洲精品在线观看二区| 亚洲一级一片aⅴ在线观看| 97热精品久久久久久| 麻豆av噜噜一区二区三区| 国产激情偷乱视频一区二区| 精品不卡国产一区二区三区| 久久国产乱子免费精品| 日本精品一区二区三区蜜桃| 一个人免费在线观看电影| 可以在线观看的亚洲视频| 久久精品国产亚洲av天美| 18+在线观看网站| 永久网站在线| 九九久久精品国产亚洲av麻豆| 中国国产av一级| 亚洲国产色片| 别揉我奶头 嗯啊视频| av在线亚洲专区| ponron亚洲| 五月玫瑰六月丁香| 亚洲七黄色美女视频| 久久精品国产亚洲网站| 欧美一级a爱片免费观看看| 久久鲁丝午夜福利片| 欧美日本亚洲视频在线播放| 国产一区二区亚洲精品在线观看| 亚洲精品久久国产高清桃花| 免费看日本二区| 麻豆一二三区av精品| 亚洲第一区二区三区不卡| 国产精品av视频在线免费观看| 欧美性猛交╳xxx乱大交人| 1024手机看黄色片| 大又大粗又爽又黄少妇毛片口| 成人亚洲欧美一区二区av| 日日撸夜夜添| 在现免费观看毛片| 精品久久久久久久久久免费视频| 亚洲成人av在线免费| 国产高清视频在线观看网站| 亚洲成人久久爱视频| 日韩大尺度精品在线看网址| 亚洲电影在线观看av| 丰满人妻一区二区三区视频av| 午夜福利在线观看吧| 观看免费一级毛片| 国产老妇女一区| 亚洲18禁久久av| 亚洲四区av| 国产高清三级在线| 一个人看的www免费观看视频| 久久精品国产清高在天天线| 精品欧美国产一区二区三| 又爽又黄a免费视频| 小说图片视频综合网站| 久久中文看片网| 嫩草影院入口| 亚洲在线自拍视频| 精品一区二区三区人妻视频| 国产亚洲91精品色在线| 国产精品一区www在线观看| 精品久久久久久久人妻蜜臀av| 神马国产精品三级电影在线观看| 欧美性感艳星| 99久国产av精品| 天天一区二区日本电影三级| 午夜a级毛片| 色播亚洲综合网| 级片在线观看| 夜夜看夜夜爽夜夜摸| 女生性感内裤真人,穿戴方法视频| 一进一出抽搐动态| 国产av麻豆久久久久久久| 国产女主播在线喷水免费视频网站 | 校园春色视频在线观看| 熟妇人妻久久中文字幕3abv| 亚洲乱码一区二区免费版| 天美传媒精品一区二区| 极品教师在线视频| 老熟妇仑乱视频hdxx| 欧美zozozo另类| 国产av麻豆久久久久久久| 一级a爱片免费观看的视频| 狠狠狠狠99中文字幕| 国产精品久久电影中文字幕| 波野结衣二区三区在线| 天堂动漫精品| 美女高潮的动态| 白带黄色成豆腐渣| 美女黄网站色视频| 热99re8久久精品国产| 日韩欧美在线乱码| 日本黄大片高清| 久久精品国产清高在天天线| 亚洲最大成人手机在线| 久久亚洲国产成人精品v| 精品国产三级普通话版| 插阴视频在线观看视频| 久久久精品大字幕| 99久久精品热视频| 99久久无色码亚洲精品果冻| 亚洲图色成人| 国产老妇女一区| 亚洲精品色激情综合| 麻豆国产av国片精品| 看黄色毛片网站| av中文乱码字幕在线| 少妇裸体淫交视频免费看高清| 中出人妻视频一区二区| 亚洲成人av在线免费| 久久精品综合一区二区三区| 又爽又黄无遮挡网站| 三级经典国产精品| 黑人高潮一二区| 最近的中文字幕免费完整| 伊人久久精品亚洲午夜| 日韩欧美一区二区三区在线观看| 嫩草影院入口| 可以在线观看毛片的网站| 在线观看一区二区三区| 国产精华一区二区三区| 欧美日本视频| 久久亚洲精品不卡| 精品欧美国产一区二区三| 午夜免费男女啪啪视频观看 | videossex国产| 久久亚洲国产成人精品v| 国语自产精品视频在线第100页| 久久久午夜欧美精品| 综合色丁香网| 99热6这里只有精品| 国产精品电影一区二区三区| 日韩欧美精品v在线| 好男人在线观看高清免费视频| 国产精品99久久久久久久久| 99国产极品粉嫩在线观看| 精品久久久久久久末码| 亚洲av成人精品一区久久| 亚洲精品日韩在线中文字幕 | 久久婷婷人人爽人人干人人爱| 久久久色成人| 日韩欧美在线乱码| 久久精品久久久久久噜噜老黄 | 亚洲国产高清在线一区二区三| 有码 亚洲区| 欧美3d第一页| 亚洲成人av在线免费| АⅤ资源中文在线天堂| 啦啦啦啦在线视频资源| 插逼视频在线观看| 露出奶头的视频| 日韩欧美国产在线观看| 悠悠久久av| 一区二区三区四区激情视频 | 久久精品国产鲁丝片午夜精品| 成人av一区二区三区在线看| 天美传媒精品一区二区| 老女人水多毛片| 亚洲成人久久性| 毛片一级片免费看久久久久| 免费人成在线观看视频色| 欧美日韩在线观看h| 最后的刺客免费高清国语| 欧美激情国产日韩精品一区| 免费电影在线观看免费观看| 最近在线观看免费完整版| 赤兔流量卡办理| 无遮挡黄片免费观看| 美女内射精品一级片tv| 免费av观看视频| 色综合亚洲欧美另类图片| 精品无人区乱码1区二区| 欧美不卡视频在线免费观看| 欧美性猛交黑人性爽| 人妻久久中文字幕网| 1000部很黄的大片| av在线播放精品| 青春草视频在线免费观看| 精华霜和精华液先用哪个| 精品欧美国产一区二区三| 精品福利观看| 欧美一区二区国产精品久久精品| 午夜精品国产一区二区电影 | 午夜a级毛片| 日韩 亚洲 欧美在线| 日本熟妇午夜| 变态另类丝袜制服| 中文字幕免费在线视频6| 蜜臀久久99精品久久宅男| 国产真实乱freesex| 精品少妇黑人巨大在线播放 | 真人做人爱边吃奶动态| 九九久久精品国产亚洲av麻豆| 美女xxoo啪啪120秒动态图| 我要搜黄色片| 国产v大片淫在线免费观看| 老熟妇仑乱视频hdxx| 卡戴珊不雅视频在线播放| 啦啦啦韩国在线观看视频| 亚洲成人中文字幕在线播放| 可以在线观看的亚洲视频| 91久久精品电影网| 午夜精品在线福利| 全区人妻精品视频| 一级毛片我不卡| 国产黄片美女视频| 舔av片在线| 又黄又爽又刺激的免费视频.| 一区福利在线观看| 成年免费大片在线观看| 日韩欧美 国产精品| 给我免费播放毛片高清在线观看| av福利片在线观看| 日本a在线网址| 亚洲,欧美,日韩| a级毛片免费高清观看在线播放| 国产av不卡久久| 别揉我奶头 嗯啊视频| 免费看a级黄色片| 别揉我奶头~嗯~啊~动态视频| 婷婷精品国产亚洲av| 欧美性猛交╳xxx乱大交人| 亚洲在线观看片| 国产真实伦视频高清在线观看| 亚洲最大成人av| a级一级毛片免费在线观看| 香蕉av资源在线| 男人狂女人下面高潮的视频| 国产高潮美女av| 麻豆国产av国片精品| 在线观看66精品国产| 偷拍熟女少妇极品色| 可以在线观看毛片的网站| 91麻豆精品激情在线观看国产| 欧美中文日本在线观看视频| 久久欧美精品欧美久久欧美| 国产欧美日韩一区二区精品| 99视频精品全部免费 在线| 偷拍熟女少妇极品色| 免费人成在线观看视频色| 久久精品国产亚洲av香蕉五月| 免费看日本二区| 蜜臀久久99精品久久宅男| 老司机影院成人| 日本-黄色视频高清免费观看| 久久午夜福利片| 天天一区二区日本电影三级| 国产一区二区三区在线臀色熟女| 联通29元200g的流量卡| 97超视频在线观看视频| 免费黄网站久久成人精品| 亚洲真实伦在线观看| 俺也久久电影网| 人人妻人人澡人人爽人人夜夜 | 色综合色国产| 黄色视频,在线免费观看| 亚洲在线观看片| www.色视频.com| 麻豆av噜噜一区二区三区| 一个人看的www免费观看视频| 最近视频中文字幕2019在线8| 亚洲在线观看片| 久久欧美精品欧美久久欧美| 欧美在线一区亚洲| 国产伦一二天堂av在线观看| 欧美一区二区国产精品久久精品| 亚洲在线自拍视频| 99热网站在线观看| 亚洲av电影不卡..在线观看| 99国产极品粉嫩在线观看| 国产欧美日韩一区二区精品| 国产视频一区二区在线看| 亚洲av熟女| 国产免费男女视频| 成人特级av手机在线观看| 免费观看精品视频网站| 国产私拍福利视频在线观看| 99久国产av精品| 最好的美女福利视频网| 最近2019中文字幕mv第一页| 国产精品一区二区性色av| 俺也久久电影网| 少妇高潮的动态图| 观看美女的网站| 免费人成在线观看视频色| 五月玫瑰六月丁香| 老师上课跳d突然被开到最大视频| 国内揄拍国产精品人妻在线| 91精品国产九色| 国产真实伦视频高清在线观看| 麻豆av噜噜一区二区三区| 欧美又色又爽又黄视频| 国产精品一区www在线观看| 欧美三级亚洲精品| 久久久久九九精品影院| 国产视频一区二区在线看| 国产成人a∨麻豆精品| 国语自产精品视频在线第100页| 毛片一级片免费看久久久久| 亚洲成人av在线免费| 亚洲欧美清纯卡通| 日本黄色视频三级网站网址| 亚洲欧美清纯卡通| 国内精品美女久久久久久| 简卡轻食公司| 日韩精品青青久久久久久| 麻豆乱淫一区二区| 国产 一区精品| 午夜精品国产一区二区电影 | 亚洲欧美日韩高清专用| 欧美一区二区精品小视频在线| 亚洲欧美日韩高清专用| 大型黄色视频在线免费观看| 一进一出抽搐动态| 欧美一区二区精品小视频在线| 最近2019中文字幕mv第一页| 波多野结衣高清无吗| 久久99热6这里只有精品| 国产男靠女视频免费网站| 国产精品伦人一区二区| 久久婷婷人人爽人人干人人爱| 亚洲图色成人| 亚州av有码| 午夜免费男女啪啪视频观看 | 一级毛片aaaaaa免费看小| 九九热线精品视视频播放| 亚洲成av人片在线播放无| 观看美女的网站| 亚洲色图av天堂| 男女视频在线观看网站免费| 中文亚洲av片在线观看爽| 日韩亚洲欧美综合| 91午夜精品亚洲一区二区三区| 亚洲一区高清亚洲精品| 一a级毛片在线观看| 免费黄网站久久成人精品| 久久久久久伊人网av| 精品国内亚洲2022精品成人| 男女边吃奶边做爰视频| 欧美日本视频| 男女视频在线观看网站免费| 99视频精品全部免费 在线|