• <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 ?

    Deep Learning-Based Robust Morphed Face Authentication Framework for Online Systems

    2023-12-12 15:51:16HarshMankodiyaPriyalPalkhiwalaRajeshGuptaNileshKumarJadavSudeepTanwarOsamaAlfarrajAmrTolbaMariaSimonaRaboacaandVerdesMarina
    Computers Materials&Continua 2023年10期

    Harsh Mankodiya,Priyal Palkhiwala,Rajesh Gupta,?,Nilesh Kumar Jadav,Sudeep Tanwar,Osama Alfarraj,Amr Tolba,Maria Simona Raboacaand Verdes Marina

    1Department of Computer Science and Engineering,Institute of Technology,Nirma University,Ahmedabad,382481,India

    2Computer Science Department,Community College,King Saud University,Riyadh,11437,Saudi Arabia

    3Doctoral School,University Politehnica of Bucharest,Bucharest,060042,Romania

    4Department of Hydrogen and Fuel Cell,National Research and Development Institute for Cryogenics and Isotopic Technologies,Ramnicu Valcea,240050,Romania

    5Department of Building Services,Faculty of Civil Engineering and Building Services,Technical University of Gheorghe Asachi,Iasi,700050,Romania

    ABSTRACT The amalgamation of artificial intelligence(AI)with various areas has been in the picture for the past few years.AI has enhanced the functioning of several services,such as accomplishing better budgets,automating multiple tasks,and data-driven decision-making.Conducting hassle-free polling has been one of them.However,at the onset of the coronavirus in 2020,almost all worldly affairs occurred online,and many sectors switched to digital mode.This allows attackers to find security loopholes in digital systems and exploit them for their lucrative business.This paper proposes a three-layered deep learning (DL)-based authentication framework to develop a secure online polling system.It provides a novel way to overcome security breaches during the face identity (ID) recognition and verification process for online polling systems.This verification is done by training a pixel-2-pixel Pix2pix generative adversarial network (GAN) for face image reconstruction to remove facial objects present (if any).Furthermore,image-to-image matching is done by implementing the Siamese network and comparing the result of various metrics executed on feature embeddings to obtain the outcome,thus checking the electorate credentials.

    KEYWORDS Artificial intelligence;discriminator;generator;Pix2pix GANs;Kullback-Leibler(KL)-divergence;online voting system;Siamese network

    1 Introduction

    Elections in a democratic country play a crucial role in electing a leader for the nation.Elections are conducted at particular time intervals.It is a way citizens can voice their opinions and exercise their rights.However,with the issues of maintaining credibility at the polling booths,they also raise concerns about certain malpractices.Everyone knows several election problems,such as using black money,horse trading,false promises,and violence.One such issue is duplicate or dummy votes in the online polling system[1],where the electorates can vote through another’s unique vote ID in electronic voting machines (EVMs) and favors a particular candidate.Rikwith et al.[2] presented a research work to scale up the performance of EVM using fingerprint and face recognition using the R307 sensor and Raspberry Pi boards.Though it has been proven to be much better than ballot boxes[3],it has been reported to pose some issues of bias,accessibility,and accountability.Other research has been proposed;for instance,Mondal et al.[4] optimized EVMs through deep learning (DL)-based face recognition.A Region-based Convolutional Neural Network(R-CNN)-based approach has been described for face detection and feature extraction.Moreover,references [4,5] proposed a one-step authentication framework using a biometric system.This literature shows that the face recognition scenario under occlusion due to objects such as masks,glasses,etc.,has not been considered.Such an inflection from the norm can disrupt systems,making them ineffective in detecting the faces.

    Further,there have been many reported cases of phishing attacks on online voting systems,compromising the framework of the system and,consequently,its credibility.Nisha et al.[6,7] have discussed various possibilities of firewall infringements and reported security attacks in the past years,which have questioned the authenticity of the system.In the online voting system,the chances of malpractice are quite high.One such issue is logging in by an attacker under a different registered user account by morphing their face using various props.In this way,the attacker fulfills the facematching feature and can vote by the wrong means.Thus,the system is susceptible to duplicate votes,which questions its credibility.Reference [8] discussed the advancements in morph generation and morphometric detection in face morphing attacks.It details various techniques to detect morphed images[8]effectively.Moreover,since 2020,the world has been struck by a global pandemic resulting in many deaths worldwide.It has been advised to avoid going to crowded places and mandatory to always wear masks.In such a scenario,online voting systems are considered safe for the elections.However,due to the mask,many voters can maneuver the verification process of the online polling system.Our system will work with real-time images,i.e.,images recorded from the input video stream.Hence,there is a need to employ a technique for morph detection of real-time images.

    The generative adversarial networks (GANs) framework can be used to solve the problem of morph detection through the reconstruction of the image.In the context of morph face detection,GANs can be used to generate a set of morphed images that are similar to the real images in the dataset.The GANs analyze the image and if find any change in the face feature space;then it removed the change to acquire the real image.The morphed images can then be used to train a classifier to detect morphed faces.GANs framework was first proposed by Goodfellow et al.[9] to estimate generative models via an adversarial process.It was a breakthrough discovery and gave excellent results when implemented on different datasets.Moreover,reference[9]proposed the idea of simultaneously training two models:a generative model and a discriminative model,which demonstrates a minimax two-player game.Various research works have been proposed for face recognition techniques under occlusion.In[10],the authors presented GANs with Visual Geometry Group(VGG)-19 as the base architecture developed for unmasking a masked face and regenerating a realistic image of the person.Several research works have been published by researchers across the globe presenting image-to-image translation techniques.It is observed that varied DL architectures like StarGAN [11],cycleGAN[12],DualGAN[13,14],etc.,are trained to achieve the task of image-to-image translation.One such research work[15]was carried out by Isola et al.,proposing a system for image-to-image translation using conditional GANs (cGANs).This research work uses thePix2pixGAN-based framework,specifically employed for processing an image as the input in the online voting system.The problem of image morphing in an online voting system addressed in this paper requires implementing the reconstruction process of an input image recorded from the user’s webcam.Pix2pixGAN architecture is hence used at the initial layer for this purpose.

    Furthermore,the reconstructed image is matched with the user’s image stored in the universal database,i.e.,image matching has to be performed in the system.Hence,another network architecture,the Siamese network [16],is trained.It is a neural network that compares two image input matrices and finds their similar confidence score.The greater the value of this score,the more similar the input images are.In [17],the authors proposed a method to understand the general similarity function in the Siamese network using a hybrid convolutional neural network (HybridCNN) [18] as their principal network.Table 1 shows the relative comparison of existing state-of-the-art approaches and their research gaps.This paper proposes an authentication framework to verify user credentials in light of the aforementioned issues.The framework proposed in this paper aims to check the user credentials in a three-step manner.Firstly,it matches the unique ID with the same present in the universal database.

    Table 1:A relative comparison of various state-of-the-art approaches

    Consequently,it scans the user’s live facial features,processes them using GANs trained model,and stores them encrypted in another database.Subsequently,these features are matched with the original image and uploaded with ID in the universal database to check if the user has a right to vote in the elections.Moreover,a separate encrypted database is also maintained in which the credentials of people who have already voted are stored to strengthen the performance of the polling system.Hence,in case of a security breach in which an attacker attempts to delete the electorate ids from the database,the system can still use the scanned facial features and check if similar embedding already exists in the encrypted database.Authentication fails if user embeddings already exist in this database.

    1.1 Research Novelty

    In recent years,numerous research works have been proposed to make online voting systems more secure and scalable.However,some of these frameworks[19,22]have not considered incorporating face recognition with authentication mechanisms.For example,Pooja et al.[22] proposed a secure deeplearning mechanism for fair online voting without any real-time implementation.Next,Alvi et al.[24]considered blockchain technology for digital voting systems,but there is no discussion on the One Time Password(OTP)mechanism for security purposes.To address the security gap in the literature,a new three-layered online voting system has been proposed that deploys a GANs framework for face recognition and verification.The proposed system [21] aims to enhance the security of online voting systems,which are still vulnerable to several security and data tampering attacks.The proposed system uses a dataset of original and partially covered faces to train the GANs model.The model aims to recognize and reconstruct faces under occlusion and authenticate them through a 2-step process using a trained Siamese network.The first step involves encoding the facial data,and the second step compares it with the encoded data in the database to ensure authenticity.Previous research works,such as[20]and[22],have used face detection without considering authentication for face recognition.The proposed system goes further to address this gap by incorporating GANs framework for face recognition and authentication to enhance the security of online voting systems.While some previous frameworks have partially solved issues related to privacy,authenticity,and robustness[21]in online voting systems,they are still vulnerable to security breaches and data tampering attacks.The proposed system aims to address this vulnerability by incorporating GANs framework for face recognition and verification.The system’s effectiveness and efficiency can be tested through simulations and experiments to ensure its viability in real-world scenarios.Following are the research contributions of the proposed work.

    1.2 Research Contributions

    The main contributions of this paper are as follows:

    ? We propose a systematic framework to ameliorate the security of the online voting system.It is achieved by implementing three layered-system,i.e.,GAN layer/reconstruction layer,the authentication,and the validation layer.The reconstruction of the image is consummated throughPix2pixGANs framework.Face detection is achieved on a video stream and is fed to this layer for processing and reconstruction.

    ? The framework consists of two-factor authentication,which integrates GANs framework with the credential matching process with the system’s universal database.

    ? Further,a Siamese network is utilized for authentication purposes (image id verification),wherein a similarity confidence score is generated to interpret the output and to offer proactive decision-making.

    ? At the validation layer,the Kullback Leible (KL)-divergence metric calculates the scores between the feature embeddings obtained from the Siamese network and those in the encryption database.Thus,it ensures security and privacy to the online systems using the proposed threelayered authentication.

    1.3 Organization of the Paper

    The flow of the paper is as follows.Section 2 discusses the system model and problem formulation.Section 3 explains the proposed approach for the stated problem statement.Section 4 presents the outcomes of implementing the trained model,and Section 5 concludes the paper and provides insights into future works.

    2 System Model and Problem Formulation

    2.1 System Model

    This section gives an overview of the proposed system to achieve security goals for the online voting system.Fig.1 outlines the system model for handling security breaches with face id recognition and verification.It assumed that to log in to the system,access to the webcam should be enabled.Furthermore,it is assumed that privacy terms and conditions should be accepted before registering and logging in,especially ones emphasizing that video stream data is being monitored and stored in real time.When a login attempt is made,the input credentials are verified with the credentials stored in the universal database to check if the user is registered.After a successful verification,the webcam,which is kept on from the beginning,records the input video stream and transmits it to the AI interface.The AI interface consists of 3 layers:GANs,authentication,and validation layers.At the GANs layer,face detection and reconstruction are carried out using thePix2pixGANs framework.The trainedPix2pixmodel detects any facial objects,like a face mask,glasses,etc.,present on the face and reconstructs it and consequently stores the feature vectors obtained from the resulting image in an encrypted database.A trained Siamese network is implemented in the next layer,which checks for similarity scores between the reconstructed image and the original image id from the universal database.The final layer is the validation layer,where the KL divergence score is calculated between feature embeddings obtained and stored from the previous layer.The final portal gets activated only after successfully executing the AI interface and its decision.The layers and their functioning are discussed in detail in the upcoming sections.

    Figure 1:System model to achieve security goals for the online voting system

    2.2 Problem Formulation

    With the digitization wave,almost every sector has adopted online strategies to increase business or to make services available to users at their convenience.However,traditional electronic voting systems are not flexible and have a lot of hassles.A paramount event for any democratic state is conducting elections online by developing robust and secured voting interfaces.With online polling systems,more people who cannot reach the allotted voting booths can cast their votes easily.The online voting system can provide convenience by increasing accessibility.Moreover,the automated systems prevent long queues and waiting times,making the process highly efficient.However,some security issues might exist that can disrupt the voting system.A security breach in a voting system can pose severe risks or even destroy a democratic state.Reference[6]detailed various phishing attacks that can occur in the online voting system.

    This paper has addressed a possible issue that can cause a security breach in the most crucial issue of an online voting system.The module discussed here is face id recognition and verification.Consider the following scenario:a voterVwith a unique voter idVidlog into the online system.We assume a universal database Ψ containing the existing and registered unique voter id and id imageVimage.A citizen can only cast their vote if,firstly,VidεΨ and face input imageVinpfrom the camera video stream andVimagematch with each other.(Note:It is assumed that proper facilities of webcam and internet are available to the voter.)The physical appearance of a person,especially a face,can be easily morphed to look like someone else.This can lead to fallacious registration of the vote by mismatched voter id and the actual id face.

    whereaandbare two instances from the universal database.Hence,fake voter registration is a concern caused by physically doctoring one’s appearance.This paper aims to prevent such altercation by employing image reconstruction using a DL approach.The reconstructed imagefor a person,xcan be produced,whic(h can be)compared withwith authenticated voter idTrained image verification pipelineIVgenerates a confidence score λ for image similarity.

    All the registered votes are stored in a separate database Δ.A major issue with such a two-layered approach is the possibility of re-registration of votes using the same voter id in case of an attack αpkc.An attack αpkc(primary key compromise attack)can be developed and deployed such that it queries to delete the tokenization voter idfor a specific personx,such that a new redundant vote can be recast again from the same voter id

    Hence,to solve the security issues identified above,a three-layered framework is proposed in the forthcoming section.

    3 Proposed Approach

    The problem formulation has discussed a few possible problems in implementing a classic system.Fig.2 presents the AI layer of the proposed system model that aims to prevent shortcomings of the online electoral system.This section discusses the proposed three-layered framework to prevent physical security breaches for the approach taken into consideration.

    Figure 2:The proposed AI layer

    3.1 GAN Layer

    The GAN layer is the foremost layer in the decision layer of the proposed framework.This layer consists of a trained generative model that aims to generate a reconstructed image of the input video stream of the electorate’s face.In this subsection,we describe the generative model training,prediction,and dataset.

    Dataset Description—No publicly available dataset contains image pairs of faces with and without attached objects such as specs,masks,hats,sunglasses,etc.Hence,we created a synthetic dataset that matches the use case description.The data collection step involves obtaining face images of different people with and without facial objects attached to them.A total of 20 image pairs were collected from 5 different people.Images of the high resolution were later resized to 256×256 to reduce computational expenses.The images are captured inside a light-illuminated room of a machine learning laboratory on the university campus.A dataset is prepared such that an image pair is captured and obtained for a person.This pair contains two different images of the same person.During the dataset preparation process,we restricted the face images to only the front face,not the lateral one.Furthermore,we only considered cases with only one face per frame.Detecting multiple faces in a frame is another research area to work on in the future.Given the novelty of the research,we constrained our dataset to only clear-face images and exempted frames with blurred faces and cases with different illumination conditions.Fig.3 shows a sample image pair present in the dataset.Label image(Fig.3b)here refers to the image that the generator sub-layer is supposed to reconstruct/regenerate given an original input image(Fig.3a)with several attached facial objects,i.e.,a blue face mask and glasses.

    Dataset Preparation and Processing—The data preparation and processing step involves making the image data fit for training and testing.LetIbe the set ofPimage pairs denoted byip.

    whereopandlpare the original and label images,respectively,for thepthinstance in the dataset.Eachip=(op,lp)∈I,op,andlpwith size(h′×w′)are resized to a static shape(h×w).

    Note:Total number of color channels in the images is kept constant,i.e.,three.

    The final pre-processed datasetImis obtained,wheremdenotes the total number of instances of image pairs.Furthermore,the image pixel values having a range[0,255]are also scaled to a new range of[-1,1]using the min-max scaler.

    Generative Adversarial Networks—The generative adversarial networks are deep learning-based generative models.Such a network[9]of the model was first developed and presented by Goodfellow.A basic GAN architecture is divided into two main sub-models: generator and discriminator.The generator model learns to mimic the original training data,whereas the discriminator [25] learns to classify it as real or generated.GANs can be used in face recognition to generate realistic synthetic faces that can be used for training face recognition models.This can be useful when there is a limited amount of real face data available for training or when the data needs to be augmented to improve the robustness of the model.In an online voting system,GANs can be used to generate synthetic faces of voters,which can be used to facilitate remote voting.This can be useful when voters cannot physically attend a voting station during a pandemic or in remote areas.The GAN-generated faces can be used to verify the voter’s identity,and the system can compare the synthetic face with the real face of the voter using face recognition technology.The pre-trained generator model weights can be preserved to generate newer instances with nearly equivalent distributions as the training data.There are several GAN architectural variants.These variants are applied to different fields involving the usage of AI.This paper uses the image-to-image(I2I)translation method,which renovates/reconstructs the input face images by learning a masking operation.Pix2pixGANs[26],which employs supervised learning,is used here for the I2I translation task.However,it is important to note that there are ethical and legal considerations around using GANs in online voting systems,as there is a risk of impersonation and fraud.Therefore,robust security measures and regulations must be implemented to ensure the voting process’s integrity.Therefore,we consider an authentication mechanism with the Siamese network along with GANs,which strengthens the current voting systems.

    GANs Training—Pix2pixGAN is trained on the collected image dataset.We were motivated by the work of Isola et al.[26,27],where they constructed an efficient model.However,the training process is inspiringly carried out from[28],showing an efficient training time.The generator model and the discriminator are trained separately on a given training batch.Fig.4a shows the generator sub-layers for the GAN model.A generator here has two components:an encoder and a decoder,which makes it a U-Net[29]like architecture.Fig.4b shows the discriminator sub-layer for the model.The discriminator aims to train the generator for better image generation.For input data represented by setI,consider generator sub-modelGand discriminator sub-modelD.Also,let the fully connected GAN model be represented as ζ(G,D).Consider an original imagebe passed to the generator sub-modelG.

    Figure 4:(a)Generator sub-model and(b)Discriminator sub-model

    whereZis the image reconstructed by the generator sub-layer in a single pass,it can be observed that the image shape remains maintained.The generated imageZand the original image χ can be passed to the discriminator to classify it as real or fake.The GANs loss functionLG,Dcan be defined as follows:

    wherecrefers to the class labels of the input image,i.e.,whether the image is real or fake.Note thatchas dimensions equivalent to the discriminator’s output layer.Moreover,a separate generator loss is added to the overall model loss to train the model more concretely and have near-to-ground truth generations.Generator lossLGis defined asL1distance ofZgenerated from χ0w.r.t the real label images χl.After the GAN model is trained,the reconstructed image of the input stream is obtained asZ.Here,

    Training configuration for the GAN layer is provided in Table 2.

    Table 2:Training configuration for GAN model

    3.2 Authentication Layer

    The authentication layer aims to cross-check the reconstructed input face image with the existing unique login and image ids database.The objective of this layer is to employ a trained Siamese network[16]for the task of image verification.

    Siamese Network Training—Similar to the GAN layer,the authentication layer consists of a deep neural network that does the task of image id verification.A simple Siamese network is trained for the same purpose.It is a type of neural network architecture that is designed to compare two different inputs and determine how similar they are.It was originally proposed for signature verification but has since been applied to various tasks,such as face recognition,image retrieval,and natural language processing.A Siamese network consists of two identical sub-networks which share the same weights and architecture.The two sub-networks take in the two input samples to be compared and extract their features.The outputs of the two sub-networks are then compared using a similarity metric such as Euclidean distance or cosine similarity,which produces a similarity score that indicates how similar the two inputs are.The advantage of using a Siamese network is that it can be trained on a small dataset of pairs of inputs,rather than a large dataset of individual samples.This makes it particularly useful for tasks where labeled data is scarce,such as one-shot learning and few-shot learning.This model takes in two input images of the same shape and outputs a confidence score pertaining to their mutual similarity.A higher output probability for a given pair of images implies a greater visual similarity in them.Training configuration for the Siamese network is provided in Table 3.In addition,Fig.5 shows the network architecture constructed for image verification.We refer to reference[30]to construct and train this model.

    Figure 5:Siamese network architecture

    Table 3:Training configuration for Siamese model

    Consider a training dataset S of size s containing the images ιidand ιgento be compared and their ground truth class labels μ.It is important to note here that generated image ids are resized to a static shape of 256×256×1 and also transformed to grayscale.

    whereErefers to the Siamese feature embeddings obtained at the output of the functional layer in Fig.5.Also σ refers to the Sigmoid activation function.

    Siamese Predictions—A universal database Ψ is considered to contain tuples of unique electorate ids ω and image idsO.

    For a correctly identified and existing unique id ωj,wherej≤size,image idOj,and GANgenerated imageZare passed on to the trained Siamese network Γ.A probability scorepωjis obtained,which is used further for decision-making.

    Ifpωj≥pth,the feature embeddings for the generated imageare stored in a database,and the credential verification process proceed to its third layer,i.e.,the validation layer.

    3.3 Validation Layer

    This is the only layer where no neural network is trained and used for predictions.This layer works as the final step for electorate id verification.For the given authenticated unique ids from the second layer,feature embeddingfor the generated GAN image are stored in a database called the encryption database Φ of sizeR.

    An authenticated image with feature embeddings(ω refers to the user’s unique id)is queried through the database,and divergence values are calculated using a particular metric.The main objective of the validation layer is to obtain divergence scores between existing data of feature embeddings and new feature embeddings to check for redundancy.KL-divergence [31] is used as a metric here.A higher KL-divergence score suggests a greater difference in values of the two input distributions proving that the electorate image id has still not been registered.

    The detailed analysis of results for all three layers is carried out in the forthcoming section.

    4 Results and Discussions

    This section presents a detailed description of the performance and evaluation of the individual components of the proposed three-layered system framework.

    4.1 GANs Results

    The purpose of the generator in the proposed architecture is to extract and reconstruct the original face without external facial objects such as glasses and face masks.Pix2pix,which is a GANs-based approach described in the proposed architecture,is trained and used for the predictions.Fig.6 shows the loss curvesvs.iterations for the discriminator of the GAN model.The models were trained for 4000 iterations,and it can be observed that the overall losses do decrease at the end of the total epochs.The loss curve for the generator sub-layer of thePix2pixGAN architecture is also plotted to check its training performance.Fig.7 shows the training loss curve for the generator sub-layer.As the model learns to reconstruct the original images by masking out the facial objects,the relative divergence between the predicted training and the original images depicted by the generator loss function also decreases with the iterations.

    Figure 6:Discriminator losses for the proposed GAN model

    As the generator model gets trained gradually,the reconstructed generated output for a given input image also improves.A series of images for different facial configurations are passed through the generator at various intermediate training steps to obtain several reconstructed images.These images are juxtaposed to evaluate and visualize the iterative training of the generator sub-layer.Fig.8 shows the regeneration of original images using the generator model to obtain faces without the face mask and glasses.

    Figure 7:Training curve for the generator loss function

    4.2 Siamese Network-Based Results

    The Siamese network is used to validate the id image existing in the universal credential database.The reconstructed image by the generator of the trained GAN architecture can be compared with the id image using a trained Siamese network,as shown in Fig.8.A return of a higher probability score suggests a greater similarity between the compared input images.Fig.9 shows the training loss curve for the Siamese network.The model loss value decreases as the model trains gradually.Fig.10 shows similar images that are taken into consideration for comparison.A similarity confidence score of 0.72 was obtained when these images were passed to the proposed Siamese network.This essentially implies that the input images are highly likely to be similar and of the same person.Fig.11 depicts the dissimilar images that are taken into consideration.A confidence score of 0.02 was obtained on the trained Siamese network implying that the images are not of the same person,which stands out to be the truth.

    Figure 9:Siamese network’s training loss w.r.t total iterations

    Figure 11:Comparing original image id(a)and test image(b)using the Siamese network

    4.3 Validation Layer Results

    The third layer of the proposed framework,as discussed earlier,is the validation layer.The Siamese network embeddings for each input image are stored unanimously.The results of the authentication parameter defined by the KL-divergence are obtained.Figs.12 and 13 show the image embeddings obtained from the Siamese network.For images,Figs.10a and 10b,a KL-divergence of 1.1855 is obtained,implying that similar images have similar face embeddings hence a lesser KL-divergence value.For images in Figs.11a and 11b,KL-divergence of 50.1229 is obtained,implying that dissimilar images have far from equivalent face embeddings.

    5 Conclusion and Discussions

    This paper highlights one of the potential risks of organizing an online voting system.It confers a probable framework based on GANs model and Siamese network to avoid future security breaches.Generator and discriminator models have been trained,and their performances have been studied using loss functions.It gave at-par results with the reconstruction of facial features,as seen in the results section.The resulting metrics of the Siamese network,namely probability scores and KL-divergence,are considered for authentication and validation layers,respectively.These metrics have given promising results and shown greater KL divergence between different images,thereby successfully detecting fraud.

    In the future,we plan to expand our dataset containing all the possible image orientations.The dataset would consist of the cases where a lateral view of the face image is detected,on which the system will state an unsuccessful login attempt.This will make the system highly efficient and robust.Furthermore,the dataset can also include faces under different illumination conditions and faces under occlusion with varying objects like hair,hat,etc.The research is currently focused on proposing a novel approach to an online voting system.In future works,we plan to optimize the system with training on a diversified dataset and enhance the framework.

    Acknowledgement:The authors would also like to acknowledge all members of Sudeep Tanwar’s Research Group(ST Lab)for their support in revising this manuscript.Further,the authors would like to acknowledge editorial board members of CMC and reviewers for providing technical comments to improve overall scientific depth of the manuscript.

    Funding Statement:This work was funded by the Researchers Supporting Project Number(RSP2023R 102)King Saud University,Riyadh,Saudi Arabia.

    Author Contributions:The authors confirm contribution to the paper as follows:study conception and design:Rajesh Gupta,Nilesh Kumar Jadav,Priyal Palkhiwala,Osama Alfarraj,Sudeep Tanwar;data collection: Harsh Mankodiya,Priyal Palkhiwala,Rajesh Gupta,Nilesh Kumar Jadav,Amr Tolba;analysis and interpretation of results: Harsh Mankodiya,Sudeep Tanwar,Maria Simona Raboaca,Amr Tolba,Verdes Marina;draft manuscript preparation:Osama Alfarraj,Maria Simona Raboaca,Verdes Marina.All authors reviewed the results and approved the final version of the manuscript.

    Availability of Data and Materials:Not applicable.

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

    看免费av毛片| 俺也久久电影网| 最近在线观看免费完整版| 亚洲av一区综合| 精品一区二区免费观看| 老师上课跳d突然被开到最大视频 久久午夜综合久久蜜桃 | 精品人妻1区二区| 91九色精品人成在线观看| 精品人妻一区二区三区麻豆 | 国内精品美女久久久久久| 亚洲 欧美 日韩 在线 免费| 日本黄色片子视频| 午夜福利视频1000在线观看| 老司机午夜福利在线观看视频| 国产在线精品亚洲第一网站| 色5月婷婷丁香| 国产免费一级a男人的天堂| 久久国产精品影院| 久久这里只有精品中国| 欧美日韩中文字幕国产精品一区二区三区| 欧美性感艳星| 69人妻影院| 亚洲黑人精品在线| 欧洲精品卡2卡3卡4卡5卡区| 日本一本二区三区精品| 精品欧美国产一区二区三| 99久久精品国产亚洲精品| 成人午夜高清在线视频| 国产一区二区三区在线臀色熟女| 亚洲精品影视一区二区三区av| 午夜老司机福利剧场| 国产日本99.免费观看| 麻豆av噜噜一区二区三区| 99热精品在线国产| 国产白丝娇喘喷水9色精品| 人人妻人人澡欧美一区二区| www.999成人在线观看| 欧美xxxx黑人xx丫x性爽| 禁无遮挡网站| 国产黄a三级三级三级人| 99视频精品全部免费 在线| 亚洲,欧美,日韩| 国产伦人伦偷精品视频| 免费高清视频大片| 精品一区二区免费观看| 在线国产一区二区在线| 欧美性感艳星| 国产三级中文精品| 国产精品久久久久久久久免 | 我的老师免费观看完整版| 欧美+日韩+精品| 亚洲av成人精品一区久久| 麻豆av噜噜一区二区三区| 一区福利在线观看| 日韩欧美 国产精品| 免费无遮挡裸体视频| 久久久久久久久中文| 亚洲在线自拍视频| 亚洲在线观看片| 蜜桃亚洲精品一区二区三区| 特大巨黑吊av在线直播| 久久久久久久久久成人| 国产av在哪里看| 香蕉av资源在线| 在线免费观看的www视频| av在线天堂中文字幕| www.www免费av| 99热这里只有精品一区| 日本成人三级电影网站| 久久久成人免费电影| 国产精品久久久久久人妻精品电影| 亚洲成人久久爱视频| 真实男女啪啪啪动态图| 两性午夜刺激爽爽歪歪视频在线观看| 18禁黄网站禁片免费观看直播| 黄片小视频在线播放| 一个人观看的视频www高清免费观看| 国产精品美女特级片免费视频播放器| av专区在线播放| av福利片在线观看| 亚洲欧美日韩无卡精品| 色播亚洲综合网| 国产日本99.免费观看| 波多野结衣巨乳人妻| 日本黄大片高清| 一级a爱片免费观看的视频| 中文字幕熟女人妻在线| 国产成人aa在线观看| 国产视频内射| 91在线观看av| 2021天堂中文幕一二区在线观| www.999成人在线观看| 男人和女人高潮做爰伦理| 午夜激情福利司机影院| 97热精品久久久久久| 美女 人体艺术 gogo| 久久欧美精品欧美久久欧美| 亚洲美女搞黄在线观看 | 色av中文字幕| 在线观看午夜福利视频| 国产精品久久视频播放| 一夜夜www| 国产精品精品国产色婷婷| 久久精品综合一区二区三区| 国产一区二区三区在线臀色熟女| 亚洲精品粉嫩美女一区| 色吧在线观看| 亚洲av电影在线进入| 99久久精品国产亚洲精品| 国产伦精品一区二区三区四那| 日韩欧美 国产精品| 欧美午夜高清在线| 性插视频无遮挡在线免费观看| 熟女人妻精品中文字幕| 亚洲国产精品久久男人天堂| 成年女人看的毛片在线观看| 精品午夜福利在线看| 国产亚洲av嫩草精品影院| 国产久久久一区二区三区| 看黄色毛片网站| 成年女人看的毛片在线观看| 久久婷婷人人爽人人干人人爱| 啪啪无遮挡十八禁网站| 日本在线视频免费播放| 久久午夜亚洲精品久久| 日韩大尺度精品在线看网址| 欧美日本亚洲视频在线播放| 男女视频在线观看网站免费| 淫妇啪啪啪对白视频| 亚洲中文字幕日韩| 午夜福利欧美成人| 最近中文字幕高清免费大全6 | 亚洲人成网站在线播放欧美日韩| 性色av乱码一区二区三区2| 一进一出抽搐gif免费好疼| 一个人看视频在线观看www免费| 欧美高清成人免费视频www| 亚洲av电影不卡..在线观看| 亚洲av免费高清在线观看| 久9热在线精品视频| 久久久国产成人免费| 亚洲成a人片在线一区二区| 国产精品久久久久久久久免 | 免费在线观看成人毛片| 99久国产av精品| 丰满人妻熟妇乱又伦精品不卡| 两人在一起打扑克的视频| 别揉我奶头 嗯啊视频| 日韩欧美国产在线观看| 真人做人爱边吃奶动态| 免费无遮挡裸体视频| 国产91精品成人一区二区三区| 91在线精品国自产拍蜜月| 色尼玛亚洲综合影院| 欧美日韩国产亚洲二区| 国产欧美日韩一区二区三| 最后的刺客免费高清国语| 欧美一级a爱片免费观看看| 亚洲专区国产一区二区| 亚洲,欧美精品.| 久久午夜亚洲精品久久| 午夜精品一区二区三区免费看| av在线蜜桃| 琪琪午夜伦伦电影理论片6080| 精品熟女少妇八av免费久了| 国产免费男女视频| 一进一出好大好爽视频| 欧美丝袜亚洲另类 | 国产欧美日韩一区二区精品| 亚洲午夜理论影院| 久久精品人妻少妇| 欧美成人a在线观看| 如何舔出高潮| 亚洲专区国产一区二区| 国产一区二区在线av高清观看| 国产淫片久久久久久久久 | 在线观看av片永久免费下载| 综合色av麻豆| aaaaa片日本免费| 黄色配什么色好看| 国产精品免费一区二区三区在线| 成年女人毛片免费观看观看9| 91在线观看av| 国产免费av片在线观看野外av| 老鸭窝网址在线观看| 国产色爽女视频免费观看| 国产极品精品免费视频能看的| 亚洲av五月六月丁香网| 18禁黄网站禁片免费观看直播| 男女之事视频高清在线观看| 久久久精品大字幕| 一级黄色大片毛片| 天堂网av新在线| 国产精品久久久久久人妻精品电影| 成人欧美大片| 无人区码免费观看不卡| 日韩中字成人| 亚洲人成网站在线播| 69av精品久久久久久| 成人国产一区最新在线观看| 在线国产一区二区在线| 美女被艹到高潮喷水动态| 99热6这里只有精品| 中文字幕人成人乱码亚洲影| 久久久久久九九精品二区国产| 一级作爱视频免费观看| 国模一区二区三区四区视频| 悠悠久久av| 亚洲成人免费电影在线观看| 韩国av一区二区三区四区| 免费人成在线观看视频色| 美女 人体艺术 gogo| 99久久无色码亚洲精品果冻| 欧美黑人巨大hd| 一夜夜www| 丝袜美腿在线中文| 听说在线观看完整版免费高清| av在线天堂中文字幕| 91午夜精品亚洲一区二区三区 | 18+在线观看网站| 一二三四社区在线视频社区8| 最好的美女福利视频网| 国产精品一区二区性色av| 日日摸夜夜添夜夜添av毛片 | 免费搜索国产男女视频| 午夜福利免费观看在线| 精品久久久久久,| 国产精品久久久久久久久免 | 国产野战对白在线观看| 欧美激情国产日韩精品一区| 国产成人aa在线观看| 成年人黄色毛片网站| 久久精品影院6| 女生性感内裤真人,穿戴方法视频| 久久热精品热| .国产精品久久| 麻豆一二三区av精品| 极品教师在线视频| 日韩高清综合在线| 狠狠狠狠99中文字幕| 国产乱人视频| 欧美高清成人免费视频www| 麻豆国产97在线/欧美| 美女免费视频网站| 国产探花极品一区二区| 中文资源天堂在线| 久久午夜福利片| 久久久国产成人免费| 国产一区二区在线av高清观看| 免费av毛片视频| 在线观看免费视频日本深夜| 精品熟女少妇八av免费久了| 国产国拍精品亚洲av在线观看| 国产精品久久久久久亚洲av鲁大| 国产伦一二天堂av在线观看| 人妻制服诱惑在线中文字幕| 尤物成人国产欧美一区二区三区| 亚洲欧美日韩高清专用| 欧美黑人巨大hd| 午夜免费激情av| av在线天堂中文字幕| 成人三级黄色视频| 国产伦一二天堂av在线观看| 日韩大尺度精品在线看网址| 亚洲最大成人手机在线| 国产精品爽爽va在线观看网站| 国产男靠女视频免费网站| 3wmmmm亚洲av在线观看| 日韩中字成人| 欧美黄色片欧美黄色片| 波野结衣二区三区在线| 国产熟女xx| 亚洲av中文字字幕乱码综合| 波多野结衣巨乳人妻| 国产在视频线在精品| a级毛片免费高清观看在线播放| 亚洲自偷自拍三级| 亚洲成人中文字幕在线播放| 久久国产乱子伦精品免费另类| 成人欧美大片| 亚洲一区二区三区不卡视频| 乱码一卡2卡4卡精品| 亚洲一区高清亚洲精品| 禁无遮挡网站| 亚洲精华国产精华精| 热99在线观看视频| 日韩欧美在线二视频| 亚洲一区高清亚洲精品| 舔av片在线| 人妻丰满熟妇av一区二区三区| 九色国产91popny在线| 高清在线国产一区| 国产淫片久久久久久久久 | 少妇丰满av| 午夜福利18| 久久精品综合一区二区三区| 亚洲乱码一区二区免费版| 少妇被粗大猛烈的视频| 日韩国内少妇激情av| 日本免费一区二区三区高清不卡| 午夜福利18| 两性午夜刺激爽爽歪歪视频在线观看| 高潮久久久久久久久久久不卡| 少妇高潮的动态图| 亚洲第一区二区三区不卡| 老女人水多毛片| 久久久久久久久中文| 全区人妻精品视频| 美女大奶头视频| 哪里可以看免费的av片| 久久久久亚洲av毛片大全| 哪里可以看免费的av片| 国产精品久久久久久人妻精品电影| 久久午夜亚洲精品久久| 久久午夜福利片| 成人鲁丝片一二三区免费| 欧美潮喷喷水| 99久国产av精品| 赤兔流量卡办理| 18禁黄网站禁片午夜丰满| 无遮挡黄片免费观看| 天堂√8在线中文| 99热只有精品国产| 欧美激情在线99| 久久性视频一级片| 麻豆国产av国片精品| 日本一二三区视频观看| 欧美日韩中文字幕国产精品一区二区三区| 日韩亚洲欧美综合| 91在线精品国自产拍蜜月| 久久久久九九精品影院| 国产精品久久电影中文字幕| 欧美xxxx性猛交bbbb| 久久久色成人| 成年免费大片在线观看| 欧美国产日韩亚洲一区| 日日摸夜夜添夜夜添av毛片 | 久久久久久国产a免费观看| 中亚洲国语对白在线视频| www.www免费av| 国产v大片淫在线免费观看| 成人一区二区视频在线观看| 日韩欧美在线乱码| 日韩精品中文字幕看吧| 国产精品永久免费网站| 校园春色视频在线观看| 色综合婷婷激情| 国产精品综合久久久久久久免费| 亚洲片人在线观看| 欧美一区二区国产精品久久精品| 免费电影在线观看免费观看| 亚洲狠狠婷婷综合久久图片| 精品久久国产蜜桃| 国产精品一及| 国产欧美日韩精品亚洲av| 免费看a级黄色片| 欧美午夜高清在线| 99国产综合亚洲精品| 夜夜夜夜夜久久久久| 最近最新免费中文字幕在线| 1024手机看黄色片| 日本三级黄在线观看| 欧美成人a在线观看| 免费在线观看亚洲国产| 亚洲男人的天堂狠狠| 精品国内亚洲2022精品成人| 久久午夜亚洲精品久久| 免费看日本二区| 亚洲不卡免费看| 激情在线观看视频在线高清| 美女黄网站色视频| 变态另类丝袜制服| 99国产精品一区二区蜜桃av| 国产91精品成人一区二区三区| 性插视频无遮挡在线免费观看| 国产成人欧美在线观看| 女人被狂操c到高潮| 一区二区三区激情视频| 嫩草影院精品99| 久久国产精品人妻蜜桃| 精品人妻视频免费看| 九色成人免费人妻av| 精品99又大又爽又粗少妇毛片 | 日韩欧美在线二视频| 在线观看免费视频日本深夜| 国产成人福利小说| 99久久成人亚洲精品观看| av在线蜜桃| aaaaa片日本免费| x7x7x7水蜜桃| 黄色日韩在线| 久久精品久久久久久噜噜老黄 | 免费黄网站久久成人精品 | 国产一级毛片七仙女欲春2| 国产成年人精品一区二区| 久久久久九九精品影院| 国产精品不卡视频一区二区 | 51国产日韩欧美| 人妻制服诱惑在线中文字幕| 国产亚洲欧美在线一区二区| 亚洲人成伊人成综合网2020| 亚洲 欧美 日韩 在线 免费| 91在线精品国自产拍蜜月| 午夜日韩欧美国产| 久久精品人妻少妇| 18禁裸乳无遮挡免费网站照片| 国内精品久久久久久久电影| 欧美日韩综合久久久久久 | 每晚都被弄得嗷嗷叫到高潮| 中文字幕人成人乱码亚洲影| 在线免费观看不下载黄p国产 | 精品人妻一区二区三区麻豆 | 欧美在线一区亚洲| 麻豆成人午夜福利视频| 国语自产精品视频在线第100页| 精品福利观看| 男人舔女人下体高潮全视频| 有码 亚洲区| 别揉我奶头 嗯啊视频| 99精品在免费线老司机午夜| 精品久久久久久,| 亚洲aⅴ乱码一区二区在线播放| 久久久色成人| 亚洲不卡免费看| 搞女人的毛片| 美女高潮喷水抽搐中文字幕| 一本一本综合久久| 国产大屁股一区二区在线视频| 人妻夜夜爽99麻豆av| 波多野结衣巨乳人妻| 精品久久久久久久久久久久久| 两性午夜刺激爽爽歪歪视频在线观看| 国产精品久久久久久精品电影| 最近最新免费中文字幕在线| 久久精品影院6| 嫁个100分男人电影在线观看| 在线免费观看不下载黄p国产 | 免费高清视频大片| 我要搜黄色片| 国产成人啪精品午夜网站| 熟女电影av网| 亚洲内射少妇av| 一级黄片播放器| 亚洲五月婷婷丁香| 亚洲国产日韩欧美精品在线观看| 亚洲欧美日韩东京热| 国产精品一及| 免费看a级黄色片| 日日干狠狠操夜夜爽| 国产欧美日韩精品一区二区| 国产三级在线视频| 极品教师在线免费播放| 国产真实乱freesex| 亚洲欧美精品综合久久99| 午夜福利在线在线| 九色成人免费人妻av| 能在线免费观看的黄片| 色综合婷婷激情| 我的老师免费观看完整版| 精华霜和精华液先用哪个| 淫妇啪啪啪对白视频| 久久久久亚洲av毛片大全| 日韩欧美精品v在线| 亚洲18禁久久av| 日本三级黄在线观看| 91久久精品电影网| 日本精品一区二区三区蜜桃| 欧美黄色片欧美黄色片| 男人狂女人下面高潮的视频| 国产在线精品亚洲第一网站| 久久久久亚洲av毛片大全| 制服丝袜大香蕉在线| 婷婷丁香在线五月| 国产精品99久久久久久久久| 亚洲av.av天堂| 国内毛片毛片毛片毛片毛片| 亚洲第一欧美日韩一区二区三区| 国产亚洲精品av在线| 九九久久精品国产亚洲av麻豆| 淫秽高清视频在线观看| 免费人成在线观看视频色| 欧美+日韩+精品| 国产蜜桃级精品一区二区三区| 亚洲黑人精品在线| 757午夜福利合集在线观看| av天堂在线播放| 久久6这里有精品| 老师上课跳d突然被开到最大视频 久久午夜综合久久蜜桃 | 日韩中文字幕欧美一区二区| 亚洲av一区综合| 午夜精品在线福利| 国产综合懂色| 在线观看一区二区三区| 可以在线观看的亚洲视频| 观看免费一级毛片| 国产探花极品一区二区| 宅男免费午夜| 成人性生交大片免费视频hd| 国产亚洲精品久久久久久毛片| 五月玫瑰六月丁香| aaaaa片日本免费| 午夜免费成人在线视频| 亚洲成a人片在线一区二区| 国产高潮美女av| 男女视频在线观看网站免费| 中文在线观看免费www的网站| 12—13女人毛片做爰片一| 嫩草影院新地址| 国产亚洲av嫩草精品影院| 国产精品不卡视频一区二区 | 欧美日本亚洲视频在线播放| 最近最新免费中文字幕在线| 日本撒尿小便嘘嘘汇集6| 变态另类成人亚洲欧美熟女| 90打野战视频偷拍视频| 日本 欧美在线| 久久人妻av系列| 欧美+亚洲+日韩+国产| 精品乱码久久久久久99久播| 99久久成人亚洲精品观看| 国产精品免费一区二区三区在线| 蜜桃久久精品国产亚洲av| 两性午夜刺激爽爽歪歪视频在线观看| 国产精品爽爽va在线观看网站| 18禁裸乳无遮挡免费网站照片| 国产精品人妻久久久久久| 国内揄拍国产精品人妻在线| 国产av在哪里看| 午夜免费激情av| 欧美日韩瑟瑟在线播放| 九九热线精品视视频播放| 免费看日本二区| 国产极品精品免费视频能看的| 简卡轻食公司| 国产精品不卡视频一区二区 | 免费电影在线观看免费观看| 国产 一区 欧美 日韩| 亚洲欧美日韩无卡精品| 亚洲aⅴ乱码一区二区在线播放| av专区在线播放| 波野结衣二区三区在线| 美女大奶头视频| 久久久久久久精品吃奶| 国产在线男女| 变态另类成人亚洲欧美熟女| 国产精品日韩av在线免费观看| 久久九九热精品免费| 日本熟妇午夜| 国产精品久久久久久精品电影| 一本久久中文字幕| 欧美+日韩+精品| 久久久久久久午夜电影| 日本一本二区三区精品| 成人国产一区最新在线观看| 久久久久久国产a免费观看| 亚洲成a人片在线一区二区| 99精品在免费线老司机午夜| 国产一区二区三区视频了| 麻豆成人午夜福利视频| 最近视频中文字幕2019在线8| 波多野结衣高清作品| 日韩欧美国产在线观看| 特级一级黄色大片| 热99在线观看视频| 久久草成人影院| 亚洲美女视频黄频| 在线看三级毛片| 国产精品亚洲美女久久久| 免费电影在线观看免费观看| 好男人在线观看高清免费视频| 白带黄色成豆腐渣| 亚洲熟妇中文字幕五十中出| 草草在线视频免费看| 色视频www国产| 在线观看av片永久免费下载| 99国产极品粉嫩在线观看| a级毛片免费高清观看在线播放| 色尼玛亚洲综合影院| 亚洲av不卡在线观看| 麻豆久久精品国产亚洲av| 中亚洲国语对白在线视频| 亚洲人成伊人成综合网2020| 91字幕亚洲| 99热这里只有是精品在线观看 | av国产免费在线观看| 国产精品日韩av在线免费观看| 亚洲av不卡在线观看| 欧美在线黄色| 两人在一起打扑克的视频| 好男人电影高清在线观看| 国模一区二区三区四区视频| 日本黄色片子视频| av福利片在线观看| 国产伦人伦偷精品视频| 亚洲男人的天堂狠狠| 亚洲,欧美精品.| 欧美高清成人免费视频www| 亚洲黑人精品在线| 国产精品亚洲美女久久久| 日本a在线网址| 三级男女做爰猛烈吃奶摸视频| 欧美成狂野欧美在线观看| 日韩欧美国产一区二区入口| 亚洲黑人精品在线| av在线蜜桃| 男人狂女人下面高潮的视频| 高清日韩中文字幕在线| 内地一区二区视频在线| 国产国拍精品亚洲av在线观看| 免费一级毛片在线播放高清视频| 在线观看美女被高潮喷水网站 | 夜夜看夜夜爽夜夜摸| 午夜激情欧美在线| 久久精品夜夜夜夜夜久久蜜豆| 最后的刺客免费高清国语|