Authors: Abhishek Mittal, Pravneet Kaur, Dr. Ashish Oberoi
Certificate: View Certificate
The face spoof detection is the approach which can detect spoofed face. The face spoof detection methods has various phases which include pre-processing, feature extraction and classification. The classification algorithm can classify into two classes which are spoofed or not spoofed. The KNN approach is used previously with the GLCM algorithm for the face spoof detection which give low accuracy. In this research work, the hybrid classification method is proposed which is the combination of random forest, k nearest neighbour and SVM Classifiers. The simulation outcomes depict that the introduced method performs more efficiently in comparison with the conventional techniques with regard to accuracy.
With the maturation of computer vision technology and the advancement of GPU-based computing platforms, the development of person recognition methodologies gradually moves its target towards practice-oriented applications. The popularity of the application of biometric identification methods for example fingerprint, iris, finger vein, etc., has touched new heights in the last few decades. Notably, face biometrics takes advantage of its security and convenience to be comprehensively used in diverse fields and is integrated at various stages , e.g., login, authentication and payment, which are closely associated with individual strategic interests. However, with the development of social networks it becomes easier to get pictures or videos of any person making it possible to spoof the face authentication system. These behaviours are equally termed as face spoofing attacks. And they include three modes: face photo attack, replay attack, and 3D mask attack. Face spoofing attacks limit the application of face recognition system as well as increase its susceptibility with respect to security concerns. This brings about one of the daunting issues: given an image or video captured with a camera that contains human faces, how to extract different features and how to differentiate live faces and spoofing face effectively in a fruitful manner .
A. General Process of Face Recognition or Classification
The face recognition model consists of following steps:
B. Face Spoof Detection Techniques
Based on the different types of cues used in face spoof detection, the taxonomy of published methods includes four main types: (i) motion-based methods, (iii) texture-based methods, (iii) methods based on image quality analysis, and (iv) depth-based methods.
The detailed description of all these methods is given below:
3. Texture-Based Methods: These methods assume that the use of multiple spoofing mediums results in different surface reflectance and shape distortion, leading to a difference in texture between live and spoofed facial images. Texture-based methods extract image artifacts in spoof face images to counter both the printed photo and replayed video attacks. Comparted to motion-based methods, texture-based methods require only one image to detect the spoof. Nevertheless, a good generalization ability for a range of facial expressions, posture and spoofing methods is required when collecting training data from a few subjects under constrained circumstances . Thus, the texture features are combined with the image quality features. Consequently, the performance of face spoof detection goes better. The two popular techniques in this category are described as follow:
a. LBP: Various texture descriptors are applied to detect face spoofing; However, first preference is given to LBP. It is a grayscale illumination-invariant texture coding method in which each pixel is assigned a label after comparing it with its neighbours and the result is rendered in a binary number. All parameters of a local binary pattern exist in terms of number of neighbours, neighbourhood scope, and coding strategy. After that, the final computed label is organized in the histogram to determine the texture of the entire image or even image path.
b. Histograms of oriented gradients (HOG): This is another texture descriptor in which variations in gradient orientation are depicted in different parts of the image in an illumination-invariant style . The magnitude of gradients in different orientations is expressed in cells that are subsequently integrated into blocks. Finally, bins, cells, and blocks are normalized to compile the final feature vector.
4. Depth-Based Methods: In these techniques detailed information about the face is estimated for the discrimination of a live three-dimensional face from a spoofing face that is presented on 2D planar media. These technologies include defocusing technology, NIR sensors and light field cameras. The depth attribute is used to effectively detect printed photo and video replay attacks. In contrast, 3D depth analysis procedures have been designed to estimate facial 3D depth information in some studies . An approach based on optical flow field has been suggested for the analysis of the difference in optical flow field between a planar object and a 3D face. Another study deployed geometric invariants based on a set of facial landmarks to detect replay attacks. But, depth-based techniques require multiple frames or depth measuring equipment to estimate depth information, which can result in increased system costs.
II. LITERATURE REVIEW
A. Face Spoof Detection using Deep Learning
Polash Kumar Das, et.al (2019) suggested an approach in which the handcrafted attributes were integrated with DNN (deep neural network) attributes for constructing the discriminant face spoofing detection . LBP (Local Binary Patterns) descriptor was implemented to analyze the information related to attributes from the brightness and the chrominance channels. A technique planned on the basis of pre-trained CNN (convolutional neural network) called VGG-16 was put forward via static features for recognizing the video and printed 2D photo attacks. The suggested approach was effective for recognizing the real and spoofed images feature.
Xun Zhu, et.al (2021) focused on developing and training a CEM-RCNN (Contour Enhanced Mask- Region-based Convolution Neural Network) algorithm in order to detect the face spoofing . This algorithm employed the contour objectness for detecting the SMCs (spoofing medium contours). The experimental outcomes indicted that the developed algorithm was suitable to recognize the face images having SMCs and more efficient in contrast to the existing techniques.
Abdulkadir ?engür, et.al (2018) introduced a mechanism on the basis of TL (transfer learning) for which pre-trained CNN (convolutional neural network) model was deployed . This mechanism emphasized on investigating diverse deep attributes and comparing them on a common ground while detecting face liveness in videos. NUAA and CASIA-FASD datasets were applied in the experimentation. The results of experiments demonstrated that the introduced mechanism had generated optimal outcomes as compared to other schemes. Shilpa Garg, et.al (2020) formulated a robust and effectual method named DeBNet with the objective of detecting face liveness . The multilayer deep network was exploited to attain effective outcomes for diverse percentage of training set of images. The deep attributes were extracted and the face images were classified as real and spoofed using this method. An analysis was conducted on NUAA data set of images. The experimental results exhibited that the formulated method offered the accuracy of 99% for detecting the face liveness and HTER value of 0.31 on the utilized dataset.
Abdelrahamn Ashraf Mohamed, et.al (2021) investigated a DL (deep learning) method known as sequential CNN (convolution Neural Network) which had two phases such to extract the features and to classify the images . The CelebA-Spoof 2020 dataset was executed to perform the experiments for identifying the faces as real and spoofed. The accuracy was considered to determine the investigated method. The investigated method was capable of attaining the accuracy up to87% and area under ROC curve around 0.535.
B. Face Spoof Detection using Hybrid Technique
Wenyun Sun, et.al (2020) established an FCN-DA-LSA (Fully Convolutional Network with Domain Adaptation and Lossless Size Adaptation) technique for detecting the face spoofing . A LSA pre-processor and FCN based pixel-level classification algorithm, whose embedding was done with a DA layer, were comprised in this technique. The FCN (Fully Convolutional Network) algorithm aimed to deploy the basic properties of face spoof distortion. The generalization was enhanced across diverse domains via DA (domain adaptation). LSA was applied for preserving high-frequent spoof clues occurred due to the face recapturing procedure. The results depicted the supremacy of the established technique over the traditional methods for detecting the face spoofing.
Raden Budiarto Hadiprakoso, et.al (2020) projected a combined technique detecting face liveness and CNN (Convolutional Neural Network) classification algorithm . Two modules namely blinking eye module to compute eye openness and lip movement, and the CCN classification module were involved in this technique. A public dataset was exploited to train the projected technique. A simple facial recognition application made the deployment of this integrated technique on Android platform. According to results, the projected technique was adaptable for recognizing several facial spoof attacks.
Junqin He, et.al (2019) intended a methodology on the basis of integrating distinct color space models . This methodology concentrated on converting the colored image into YCbCr and Luv color space model so that attributes of LBP (Local Binary Patterns) were extracted and transforming the RGB image into HSV color space model with the objective of extracting CM (Color Moment) features later on.
At last, the extracted features were cascaded into SVM (Support Vector Machine) in order to classify the decision. Replay-Attack and CASIA-FASD datasets were utilized to compare the intended methodology against the traditional techniques. The experimental outcomes revealed the superiority of the intended methodology and its discrimination ability was found greater.
Xiaofeng Qu, et.al (2019) designed a shallow CNN-LE (shallow convolutional neural network with laplacian embedding) to detect the face spoofing . The face liveness was detected through diverse attributes in accurate way. Initially, four layers were comprised in the shallow CNN that led to enhance its speed. The DT-CWT (dual tree-complex wavelet transform) was assisted in extracting the dynamic texture features. Subsequently, the designed model employed these features which were integrated with the depth features. Eventually, LE was adopted for maintaining the inter-class discrimination and penalizing the distance of intra-class. The designed model attained more discriminative attributes when it was embedded with laplacian loss with the softmax loss. These attributes were useful for detect face anti-spoofing. The CASIA FASD, Replay attack and MSU USSA databases were applied to carry out the experiments. The experimental results confirmed that the designed model performed more successfully in contrast to others while detecting face anti-spoofing.
C. Face Spoof Detection using Image Processing
Graham Desmon Simanjuntak, et.al (2019) presented a method to detect a face spoofing on the basis of color distortion analysis for capturing the chromatic aberration from a face image . The color distortion was analyzed for extracting the color moment and ranked histogram attributes that resulted in creating 116 feature vector. Thereafter, PCA (Principal Component Analysis) employed these feature vectors for mitigating the dimensionality. The face images were classified as spoof or real using NB (Naïve Bayes) on the principal components. The experimental outcomes validated that the presented method offered the TPR (True Positive Rate) up to 97.4% in comparison with the existing methods.
Shan Jia, et.al (2021) suggested a new anti-spoofing technique on the basis of MC_FBC (factorized bilinear coding of multiple color channels) in order to learn the way for differentiating the real images from the spoofed ones . The discriminative and fusing complementary information was extracted from RGB and YCbCr spaces to build a principled solution for detecting the 3D (three dimensional) face spoofing. According to the experimental outcomes, the suggested technique was more effective as compared to the existing technique achieves under different scenarios.
Mayank Yadav, et.al (2018) constructed a method of detect face spoofing so that the faces were classified as real and spoofed . The KNN (K-Nearest Neighbor) algorithm was implemented to classify the faces using the approximate equal classification methods. The results of the constructed methods were analyzed with regard to accuracy and execution time. The experimental results revealed that the constructed method was assisted in maximizing the accuracy and mitigating the execution time.
Patrick P. K. Chan, et.al (2018) developed a technique to detect the face liveness with flash against 2D (two dimensional) spoofing attack . The flash was helpful for enhancing the process to different the authentic users from the malicious ones as well as alleviating the impact of the environmental factors. The information related to the images was captured using 4 texture and 2D structure descriptors. The cost to install this flash was lower and there was not any necessity of user cooperation. The experiments were conducted on the dataset in which 50 subjects were included. The experimental results indicated that the developed technique was more applicable in comparison with others in different environmental scenarios and proved more accurate and robust while detecting the face spoofing.
Vanitha A., et.al (2018) recommended an efficient CCMF (Color based Chromatic Moment Features) algorithm to detect the face spoof . This algorithm was consisted of two modules. At first, Viola Jones technique was implemented to detect the face. After that, the recommended algorithm was deployed to detect the face liveliness. NUAA, MSU MFSD datasets were applied to test this algorithm with regard to reliability. The experimental results revealed that the recommended algorithm yielded superior accuracy around 91.75% as compared to the conventional methods.
III. RESEARCH METHODOLOGY
The methods of detecting face spoof detection are applicable to detect the input as spoofed or normal. This work projected a system to detect the face spoof in diverse stages which are defined as:
These attributes are discussed as:-
This attribute is responsible for placing relatively high weights on the elements which are different from the average value of P(i,j) .
4. Classification of Data: The final stage is to develop a model for detecting the face spoof. The dataset is divided in two sets. The training set is large as it deploys 60% of the data and rest of the data is utilized in the testing set to perform the classification. KNN (K-Nearest Neighbour) is a classification algorithm planed on the basis of instance. The similar or similarity functions are utilized at individual level for relating the unknown samples to the unknowns in this algorithm. The learning process of this algorithm is slow. Moreover, the formulation and analysis of this algorithm is done at the same time. The k-nearest centers are required for assigning the larger part of class to the unspecified case. Simplicity is the major factor of this algorithm. The majority vote and its k neighbours play a significant role in this algorithm. This algorithm is a suitable for performing classification and regression. RF (Random Forest) is a ML (Machine Learning) algorithm which provides flexibility. The algorithmic tree predictors are integrated in this algorithm. RF generates optimal outcomes in all of the scenarios. Diverse kind of data can be handled using this algorithm. This algorithm allows the development of numerous trees. The promising outcomes are obtained with the integration of these trees in terms of precision. The major task of ML is to perform the classification. This work presents hyperparameters which are similar to DTs (decision trees) or bagging algorithms. The random trees are overlapped in this algorithm and their analysis can be conducted easily. To illustrate, seven random trees provide information related to some variables. 4 trees are agreed and rest are not. The SVM (Support Vector Machine) is integrated in voting procedure so that the classification can be executed. This ML algorithm is planned according to the majority of voting. In Random Forest, a random subset of attributes, available on the dataset, provides outcomes with higher accuracy. The output acquired from the RF, K-Nearest Neighbour and SVM is utilized in voting for input to vote amid the two classification algorithms and for generating optimal results.
IV. RESULT AND DISCUSSION
MATLAB tool assists in carrying out the mathematical complex computations. This tool makes the utilization of simplified C as the programming language. There are a number of inbuilt toolboxes such as mathematical toolbox, drag and drop based Graphical User Interface, Image processing, NNs etc. are comprised in this tool. The algorithms are exploited, graphs are plotted and user interfaces are designed using MATLAB. The results of the proposed model is analyzed in terms of accuracy, precision and recall.
As shown in figure 3, the attributes of the test and training images are analyzed with the help of textual feature analysis algorithm. The Voting classifier is applied which can classify the best match which is shown in the form of matched image.
Figure 4 illustrates the analysis of the performance of KNN (K-Nearest Neighbor) classification and hybrid classifier concerning precision, recall and accuracy. This demonstrates that the hybrid classifier provides higher values of all three metrics in comparison with KNN.
The Machine learning algorithm focuses on extracting the novel values from the earlier experiences. The training set is executed to carry out the segmentation, analyze the attributes and performs the classification. The test set is available in the form of an image whose identification is required for the attendance system. This research work utilizes the hybrid classifier for classifying the faces. The GLCM (grey level co-occurrence matrix) is implemented to analyze the texture attributes so that the faces can be classified. The simulation outcomes indicate that the introduced technique enhances the results up to 10%.
 Zinelabidine Boulkenafet, Jukka Komulainen and Abdenour Hadid, “Face Spoofing Detection Using Colour Texture Analysis”, 2018, IEEE Transactions on Information Forensics and Security, Vol. 45, No. 21, PP. 489-495  Di Wen, Hu Han, Anil K. Jain, “Face Spoof Detection with Image Distortion Analysis”, 2015, IEEE Transactions on Information Forensics and Security, Vol. 10, No. 4, PP.746 - 761  Shervin RahimzadehArashloo, Josef Kittler, William Christmas, “Face Spoofing Detection Based on Multiple Descriptor Fusion Using Multiscale Dynamic Binarized Statistical Image Features”, 2015, IEEE Transactions on Information Forensics and Security, Vol. 10, No. 11, PP. 2396 – 2407  Allan Pinto, HelioPedrini, William Robson Schwartz, Anderson Rocha, “Face Spoofing Detection Through Visual Codebooks of Spectral Temporal Cubes”, 2015, IEEE Transactions on Image Processing, Vol. 24, No. 12, PP. 4726 – 4740  Keyurkumar Patel, Hu Han, Anil K. Jain, “Secure Face Unlock: Spoof Detection on Smartphones”, 2016, IEEE Transactions on Information Forensics and Security, Vol. 11, No. 10, PP. 2268 – 2283  Abdulkadir ?engür, Zahid Akhtar, YamanAkbulut, Sami Ekici, Ümit Budak, “Deep Feature Extraction for Face Liveness Detection”, 2018, International Conference on Artificial Intelligence and Data Processing (IDAP), Vol. 35, No. 11, PP. 6373-6381  R. Raghavendra, Kiran B. Raja, Christoph Busch, “Presentation Attack Detection for Face Recognition Using Light Field Camera”, 2015, IEEE Transactions on Image Processing, Vol. 24, No. 3, PP. 1060 – 1075  Xiao Song, Xu Zhao, Tianwei Lin, “Face spoofing detection by fusing binocular depth and spatial pyramid coding micro-texture features”, 2017, IEEE International Conference on Image Processing (ICIP), Vol. 55, No. 13, PP. 4648-4654  Hoai Phuong Nguyen, Florent Retraint, Frédéric Morain-Nicolier, Agnès Delahaies, “Face spoofing attack detection based on the behavior of noises”, 2016, IEEE Global Conference on Signal and Information Processing (GlobalSIP), Vol. 32, No. 9, PP.7582-7589  HeniEndahUtami, Hertog Nugroho, “Face Spoof Detection by Motion Analysis on the Whole Video Frames”, 2017, 5th International Conference on Instrumentation, Communications, Information Technology, and Biomedical Engineering (ICICI-BME), Vol. 21, No. 7, PP. 8993-9001  Wonjun Kim, Sungjoo Suh, Jae-Joon Han, “Face Liveness Detection from a Single Image via Diffusion Speed Model”, 2015, IEEE Transactions on Image Processing, Volume: 24, Issue: 8  A. Fernández, J.L. Carús, R. Usamentiaga, R. Casado, “Face Recognition and Spoofing Detection System Adapted to Visually-Impaired People”, 2016, IEEE Latin America Transactions, Volume: 14, Issue: 2  Polash Kumar Das, Bin Hu, Chang Liu, Kaixin Cui, Prabhat Ranjan, Gang Xiong, “A New Approach for Face Anti-Spoofing Using Handcrafted and Deep Network Features”, 2019, IEEE International Conference on Service Operations and Logistics, and Informatics (SOLI)  Xun Zhu, Sheng Li, Xinpeng Zhang, Haoliang Li, Alex C. Kot, “Detection of Spoofing Medium Contours for Face Anti-Spoofing”, 2021, IEEE Transactions on Circuits and Systems for Video Technology  Abdulkadir ?engür, Zahid Akhtar, YamanAkbulut, Sami Ekici, Ümit Budak, “Deep Feature Extraction for Face Liveness Detection”, 2018, International Conference on Artificial Intelligence and Data Processing (IDAP)  Shilpa Garg, Sumit Mittal, Pardeep Kumar, Vijay Anant Athavale, “DeBNet: Multilayer Deep Network for Liveness Detection in Face Recognition System”, 2020, 7th International Conference on Signal Processing and Integrated Networks (SPIN)  Abdelrahamn Ashraf Mohamed, Marwan Mohamed Nagah, Mohamed Gamal Abdelmonem, Mohamed Yasser Ahmed, Mahmoud El-Sahhar, Fatma Helmy Ismail, “Face Liveness Detection Using a sequential CNN technique”, 2021, IEEE 11th Annual Computing and Communication Workshop and Conference (CCWC)  Wenyun Sun, Yu Song, Haitao Zhao, Zhong Jin, “A Face Spoofing Detection Method Based on Domain Adaptation and Lossless Size Adaptation”, 2020, IEEE Access  Raden BudiartoHadiprakoso, Hermawan Setiawan, Girinoto, “Face Anti-Spoofing Using CNN Classifier & Face liveness Detection”, 2020, 3rd International Conference on Information and Communications Technology (ICOIACT)  Junqin He, Jun Luo, “Face Spoofing Detection Based on Combining Different Color Space Models”, 2019, IEEE 4th International Conference on Image, Vision and Computing (ICIVC)  Xiaofeng Qu, Jiwen Dong, SijieNiu, “shallowCNN-LE: A shallow CNN with Laplacian Embedding for face anti-spoofing”, 2019, 14th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2019)  Graham Desmon Simanjuntak, Kurniawan Nur Ramadhani, AndityaArifianto, “Face Spoofing Detection using Color Distortion Features and Principal Component Analysis”, 2019, 7th International Conference on Information and Communication Technology (ICoICT)  Shan Jia, Xin Li, Chuanbo Hu, Guodong Guo, Zhengquan Xu, “3D Face Anti-Spoofing With Factorized Bilinear Coding”, 2021, IEEE Transactions on Circuits and Systems for Video Technology  Mayank Yadav, Kunal Gupta, “Novel Technique for Face Spoof Detection in Image Processing”, 2018, Second International Conference on Intelligent Computing and Control Systems (ICICCS)  Patrick P. K. Chan, Weiwen Liu, Danni Chen, Daniel S. Yeung, Fei Zhang, Xizhao Wang, Chien-Chang Hsu, “Face Liveness Detection Using a Flash Against 2D Spoofing Attack”, 2018, IEEE Transactions on Information Forensics and Security  Vanitha A., Vaidehi V., Vasuhi S., “Liveliness Detection in Real Time Videos using Color based Chromatic Moment Feature”, 2018, International Conference on Recent Trends in Advance Computing (ICRTAC)
Copyright © 2022 Abhishek Mittal, Pravneet Kaur, Dr. Ashish Oberoi. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.