Fingerprint Liveness Detection via Learning Multi-Modal Deep Features Chengsheng Yuan1,2 , Mingyu Chen1,2 1 School of Computer and Software, Nanjing University of Information Science and Technology, Nanjing 210044 2 Engineering Research Center of Digital Forensics, Ministry of Education, Nanjing University of Information Science and Technology, Nanjing 210044 Abstract With the widespread application of fingerprint identification systems, fraudulent attacks based on forged fingerprints have gradually increased, so it is very important to distinguish the authenticity of fingerprints. Fingerprint liveness detection tech- nology was proposed to slove this problem. In order to effectively integrate advantages of existing algorithms, this paper proposes an adaptive feature optimization module filtering distinctive multi-modal features. Firstly, we extract the ROI of fin- gerprint images and unify them to the same size as subsequent input. Secondly, three convolutional neural networks(CNN)- AlexNet, VGG16 and ResNet, are trained through processed images, whose the last fully connected layer as fingerprint feature. Then genetic algorithm is used to assign different weights to extracted features through these networks, which retain distinctive parts and eliminate invalid parts. Finally, considering that the features are extracted from CNN, optimized features are input to the fully connected layer, and then fake fingerprints are identified by softmax function. Experiments show that when feature dimensions of three networks output are 512, feature optimization module proposed can improve the detection accuracy by an average of 1.0% in the 2011 Livdet database, which finds out the more different parts of features extracted by the multi-modal network, enhancing fingerprint liveness detection performance. Keywords Fingerprint Liveness Detection, CNN, Genetic Algorithm, Multi-Modal 1. Introduction Clean and undamaged Residual fingerprint fingers samples In a highly information-based modern society, peo- Mould for ple often need to use passwords for identity authentica- fingerprint Capture fingerprint collection image tion to obtain access to various accounts. Therefore, the Real fingerprint password has become direct target of numerous hack- Repair and Fake finger enhance ers [1]. News about economic losses of users caused -print film fingerprint by theft of various passwords is endless, which means this traditional identity authentication scheme has se- Spoof the fingerprint Spoof the fingerprint Fake fingerprint rious security risks. In order to solve this problem, system system many new methods have been proposed and adopted, Collaborative forged fingerprint Non-cooperative forged fingerprint such as USB KEY, SMS password, dynamic password, Figure 1: The process of making fake fingerprints. The left etc. Among them, the authentication method based on picture is collaborative, requiring the cooperation of the user; biometrics has been favored by people. Compared with the right picture is non-cooperative, where fingerprints are other identity authentication methods, this method is forged by stealing fingerprint traces left by the user. Gener- ally, collaborative methods can obtain higher-quality forged simple, fast and reliable, so it has been used widely in fingerprints. all aspects of people’s social life. For instance, lots of business units check attendance of employees through fingerprint, customs in many countries will utilize fin- gerprints to authenticate immigrants, and smartphones hand, people also can copy fake fingerprints for them- use fingerprints to authenticate their identity or achieve selves to deceive the attendance system, as shown in quick payment [2, 3, 4, 5]. Fig.1. These fake fingerprints can be made of silica gel, However, fingerprint authentication also has certain gelatin, EcoFlex, Modasil and other materials. In addi- security risks [6]. For the one hand, human fingerprints tion, with the development of deep learning technology, are easily stolen and counterfeiters can imitate user fin- generative adversarial network can also generate suffi- gerprints to achieve illegal authentication; For the other ciently realistic fingerprint images. Forged fingerprint attacks are one of the biggest threats to fingerprint iden- 2021 International Workshop on Safety & Security of Deep Learning tification systems, which greatly reduce the reliability £ ycs_nuist@163.com (C. Yuan); cmyhhhh@163.com (M. Chen) © 2021 Copyright for this paper by its authors. Use permitted under Creative of the system and put users’ private information at risk. Commons License Attribution 4.0 International (CC BY 4.0). CEUR Workshop Proceedings http://ceur-ws.org ISSN 1613-0073 CEUR Workshop Proceedings (CEUR-WS.org) Hence, as an important auxiliary algorithm of finger- print identification system to identify authentic finger- 2.2. FLD based on texture features prints, fingerprint liveness detection has become an aca- Although fingerprint texture cannot be distinguished by demic research hotspot now. human eyes, it is a common feature in fingerprint im- ages and represented by the gray distribution of center 2. Related Work pixels and neighboring pixels. Common texture feature descriptors include Local Binary Pattern(LBP) [14], Bi- Nowadays, fingerprint liveness detection(FLD) al- nary Statistical Image Feature(BSIF) [15], Local Phase gorithms are mainly divided into two categories: Quantization(LPQ) [16], Histogram of Gradient Direc- hardware-based FLD and software-based FLD [7]. The tion(HOG) [17], etc. Jhat [18] et al proposed FLD based former uses additional professional equipment to iden- on gray-level independence to verify and distinguish the tify the authenticity of fingerprint images by measuring liveness of fingerprints. Yuan [19] et al calculated pa- skin temperature, conductivity, blood pressure, blood rameters of the co-occurrence matrix as features of the oxygen and other vital signs. Although this method fingerprint image to authenticate fingerprints. can achieve great detection accuracy, expensive equip- ment is easy for illegal users to find loopholes because 2.3. FLD based on deep learning of single identification method. With image process- ing technology, the latter analyze the difference be- According to different classification tasks, deep neu- tween real and fake fingerprint images for identification. ral networks can complete complex mapping and fea- Compared with the former, the latter is more flexible, ture extraction through self-learning, which is simpler. save costs, simplify operations and minimize additional Nogueira [20] et al introduced Convolutional Neural equipment [8, 9, 10]. Therefore, this method is also the Network (CNN) technology to fingerprint liveness de- current research focus of fingerprint liveness detection. tection. They designed a random model based on CNN Existing software-based algorithms can be divided into as the feature extractor, took the preprocessed images three categories: traditional FLD, FLD based on texture as the input and obtained the best detection results at features and FLD based on deep learning. the time. While achieving better accuracy, CNN mod- els also have some shortcomings. For example, a fixed- scale input image must be used in the input layer. Al- 2.1. Traditional FLD though cropping or scaling can solve the scale problem The traditional fingerprint liveness detection algorithm well, they can easily lead to the loss of some key texture designs an descriptor to extract distinctive features be- information and reduction of image resolution, thereby tween real and fake fingerprint images through the weakening the generalization performance. To solve heuristic algorithm. FLD based on sweat hole is the ear- this problem, Yuan [21] et al proposed a scale-equalized liest proposed fingerprint liveness detection algorithm. deep convolutional neural network (DCNNISE) model While recognizing high-resolution fingerprint images, utilizing the retained subtle texture information to fur- the quality of artificial fingerprint images is often worse ther improve the detection performance of forged finger- than that of the real. On account of rougher surface, de- prints. Moreover, the confusion matrix was applied to tail of fake fingerprints is much weaker than the real. In FLD as performance indicator in the performance evalu- consequence, Moon [11] et al proposed an idea of FLD ation for the first time. Zhang [22] et al further found based on image quality, which denoises and reconstructs that CNN model used for multi-classification of natu- fingerprint images by wavelet. The noise residuals be- ral images cannot obtain good accuracy in fingerprint tween reconstructed and original images are calculated liveness detection, because it ignored the difference be- to authenticate fingerprints. When the finger presses tween natural and fingerprint images and the shallower and rotates on the sensor, real fingerprints can produce network structure cannot mine deep features of finger- better elastic deformation than the fake. In consequence, prints. Therefore, they proposed a lightweight but pow- Antonelli [12] et al proposed FLD based on fingerprint erful network structure Slim-ResCNN.At present, Agar- skin elastic deformation for the first time. In the process wal S [23] et al found that existing FLD algorithms per- of imitating fingerprints, sweat holes on the ridge of the formed well when test dataset and train dataset sample finger’s epidermis are difficult to replicate. Accordingly, distribution are the same, but result of cross-sensor fin- FLD based on sweat holes is proposed. Manivanan [13] gerprint liveness detection is bad. In order to enhance et al used a high-pass filter to extract effective sweat hole the generalization, robustness and operability of FLD al- features and a correlation filter to locate sweat holes. gorithm, they believed that the learning model need be adaptive to the data and proposed a general EaZy model. This adaptability in the context of cross-sensor datasets embodies significant advantages. AlexNet In this paper, the main contributions are summarized as follows: (1) Multimodal fingerprint feature extraction Ac- VGG16 cording to different classification tasks, CNNs can Forged complete complex mapping through self-learning Genuine and extract high-level features of image. However, Res18 Genetic Feature Selection Module due to various depth and architecture of CNNs, the characteristics of real and fake fingerprints ex- tracted by different network models are quite dif- Figure 2: A flow diagram of FLD via learning multi-modal ferent, making the classifier have stronger perfor- deep features mance. In order to integrate the excellent character- istics of multiple CNNs, this paper attempts to use multimodal neural network models to extract vari- ous fingerprint features to make the difference be- image. By narrowing the gap between model output and tween real and fake fingerprints more obvious. label, model parameters are updated until they converge to a certain limit, thereby completing the complex map- (2) Genetic optimization module Full concatenation ping from two-dimensional images to one-dimensional of multi-dimensional features has big defects. In ad- features. The high-level features extracted by the deep dition, the feature extraction method based on CNN neural network are excellent discriminativeness, show- is similar to the black box operation, which means ing amazing performance in image recognition and clas- extracted fingerprint features are not known, mak- sification. In the field of fingerprint liveness detection, ing it impossible to determine the optimization di- CNN models can also achieve good identification re- rection of features. In consequence, genetic algo- sults and become a research hotspot in this area, such rithm is innovatively utilized to optimize extracted as AlexNet, VGG16, ResNet, etc. These networks have fingerprint feature, which automatically select the many differences in the depth level of the model or the obvious distinguishing part, so as to solve the prob- width level of the convolutional layer, and the learned lem of unknown features. By imitating the genetic fingerprint features will also be multifarious. In order to processes of life, such as crossover, mutation, and make full use of the merits of various networks, this pa- selection, the optimal real and fake fingerprint fea- per concatenates features of multiple convolutional net- tures are found in the fused feature space. Based works as general features of model reducing difficulty of on the trained CNN feature extractors, initialized forged fingerprint detection. feature populations are put into the fully connected layer to calculate the fitness. Through mutation and Population Initiaztion crossover operators, excellent performance genetic Calculate Selection Crossover Mutation information from parents is inherited and new genes Fitness are generated, promoting feature evolution. Get Best Feature (3) Broad adaptability This paper carried out model training and testing on 8 Livedet Datasets (2011, 2013) [24, 25]. In order to improve the stability Figure 3: Genetic optimization module in our method and generalization of the trained model, some op- erations were used to expand the training set, in- cluding rotation, brightening and flipping. The ex- In the field of FLD, some scholars believe that to al- perimental results show that the accuracy of finger- leviate the shortage of target samples, transfer learning print liveness detection on multiple data sets is sig- needs to be applied to the model. [26] They use 1.2 mil- nificantly improved in the highest accuracy of the lion ImageNet images (source task) to pre-train the con- modal network, which proves the effectiveness and volutional network, and stochastic gradient descent is wide adaptability of our model. used to optimize error losses. After training, all trained parameters of the convolutional layer in the source task are transferred to FLD. However, we think that this ap- proach still loses part of the characteristics of the fin- 3. Methodology gerprint image, so the pre-training step should be aban- doned. We select three classic CNNs——AlexNet, VGG16 3.1. Multimodal Deep Feature Learning and ResNet, as feature extractors, hoping to get features Without expert knowledge, deep neural networks with different concerns. After training the CNN, freeze have ability to automatically learn pixel distribution of network parameters and take out the penultimate fully connected layer as extracted features. Since the different dimensions of this layer will affect the output, fewer fea- Algorithm 1 Multimodal feature weight learning tures are difficult to support the classification of true and Input: The dataset of fingerprint 𝐷; The size of false fingerprints. In addition, more features can also initial population 𝑁 ; The fitness penalty value make model training difficult and key features cannot be 𝑀; The maximum number of iterations 𝑇 ; extracted. Therefore, we selected multiple dimensions Output: In 𝑇 generation, the feature chromosome for experimentation, such as 256, 512 and 1024. with the greatest fitness; 1: Population Initizlization: Initialize 𝑁 chro- 3.2. Genetic Algorithm mosomes with a length of 𝑆, the value of which is randomly generated 0 or 1; Because excessive features will cause dimensional dis- 2: for 𝑖 = 1, ⋯ , 𝑇 do asters, it is obviously not advisable to perform fully con- 3: Mutation: According to the evolutionary nected operations for multiple features [27]. If only 4: probability, a gene of length L in chromos some of the features are selected to construct the model, 5: -ome is reinitializedor reversed randomly. the difficulty of the learning task can be reduced and the 6: Crossover: Two parents are selected from interpretability of the model can also be increased. How- 7: the population for single-point crossover ever, how to filter and process these characteristics is 8: to produce new offspring. a big problem. Traditional feature selection algorithms 9: Fitness: Extract the fingerprint feature of are proposed by researchers based on the analysis of 10: testset at the corresponding position in the feature defects. For neural networks, unascertainty of 11: chromosome, whose accuracy is used as feature extraction process masks the source of the fea- 12: the fitness. ture. Genetic algorithms can adaptively find better solu- 13: Selection: The original fitness minus pena tions from the feature space, without other selection al- 14: -lty value 𝑀 is used as the new fitness. The gorithms. Better features are highlighted after a series of 15: roulette algorithm generates survival prob biologically inspired operations, such as crossover, mu- 16: -ability of individual, which means poorly tation, and selection. 17: adapted are more likely to be eliminated. The genetic algorithm mainly includes five steps, as 18: end for shown in fig.3. Firstly, with value of 0or 1, 𝑁 chromo- somes are randomly generated as the initial population, The way where the proportion of the fitness of indi- which is the same as feature length. 0 means discarding vidual chromosomes in the population is as the evolu- the corresponding location feature value, and 1 means tion probability is called the roulette wheel algorithm. selecting. Secondly, the single-layer fully connected Due to the high accuracy of the sub-network in the layer as a classifier, the characteristics of the fingerprint model, the chromosomes with higher fitness cannot ob- image at the corresponding position of the chromosome tain an advantage under the roulette wheel algorithm. are separated for train and evaluation. Record the clas- So we take the lowest precision in the sub-network as sification accuracy of the test set as fitness. Thirdly, ev- a penalty value, and highlight the gap between chromo- ery chromosome in the existing population is selected as somes by introducing a fitness penalty value. Through the parent chromosome with evolutionary probability to roulette wheel selection, chromosomes with high fitness produce offspring. The evolution probability calculation are more likely to be retained as parental chromosomes, formula is as follows: whose structural information is passed on to the off- spring. Fourthly, on the basis of inheriting parental chro- 𝐹𝑖 − 𝑀 mosomes, offspring chromosomes will have mutations 𝑃𝑖 = 𝑁 , (1) ∑𝑖=0 (𝐹𝑖 − 𝑀) in a certain length of gene encoding. Some randomly generated or inverted gene codes replace the original where 𝐹𝑖 is the fitness of the 𝑖𝑡ℎ chromosome, 𝑁 is the part with a certain probability, which means that certain number of chromosomes, 𝑃𝑖 is the probability that the features are reselected. The mutation of gene coding 𝑖𝑡ℎ chromosome is selected, and M is the fitness penalty provides the possibility of population evolution. Finally, value. the roulette wheel selection method is utilized to select parents with greater fitness for crossover from popula- tion. Cut off the two parental chromosomes (divided into upper and lower parts) at the same position, and ex- change the upper parts to form two brand-new chromo- somes. Each chromosome inherits powerful genes from both parents. After the gene mutation and crossover of the 𝑡 𝑡ℎ generation, many offspring will be produced and combined with the parental chromosomes to form a new (small angle rotation, flip, zoom, and brightness), dataset population. The 𝑘 chromosomes with the highest fitness is enhanced to train each sub-network (15 training are selected as the (𝑡 + 1𝑡ℎ ) generation population and epochs with learning rate 0.0002). Then, freeze the net- continue to evolve until the set generation threshold is work parameters as the feature extractor of our model. reached. Choose the best-performing individual as the Next, the genetic algorithm is implemented in the fin- optimal solution. gerprint features extracted by the above-mentioned sub- network to find the parts with significant differences. We set the initial population size N to 10, the maximum 4. Experiments evolutionary generation T to 20, and the mutation prob- ability of each chromosome to 0.05. According to the 4.1. Dataset roulette algorithm to determine the probability of each In this paper, LivDet 2011 and 2013 datasets are used chromosome being selected as a parent, all parents have to conduct experiments. The entire data set includes a 50% probability of crossover operations with other par- 16470 images as the training set and 16439 images as the ents. In the selection operation, the roulette algorithm is test set. Six different sensors (Biometrika, CrossMatch, still used to eliminate individuals with lower fitness and Identix, etc.) are used for image acquisition. The fake maintain the population size N. The experimental results fingerprint data set includes 9 different materials (Body- show that the accuracy of FLD is about 1% higher than Double, EcoFlex, gelatin, latex, Silgum, WoodGlue, etc.) the highest in the sub-network after screening by ge- to make fake fingerprint images. But the scale of the netic algorithm. The three sub-networks selected in this image is different, so it is very necessary to unify the paper have the same structure as in the original paper, size of the image. This paper solves the traditional oper- except that the dimensions of the final fully connected ations that require cropping and zooming through ROI layer are changed. Table 2 shows that the accuracy of the operations.The specific information of LivDet Datasets method proposed on the LivDet2011 dataset far exceeds is shown in the following table: that of other algorithms. Although not every dataset achieves the best accuracy on the LivDet2013 data set, our method is not far from it and the average accuracy Table 1 also has competitiveness. Details of the LivDet datasets. Image Dataset Sensor train/test size 5. Conclusion Biometrika(Bio) 2000/2000 312×372 Due to the difference in the depth and architecture LivDet2011 DigitalPersona(Dig) 2004/2000 355×391 of CNN, the fingerprint features extracted by each net- Italdata(Ita) 2000/2000 640×480 work are not the same. In order to combine their ad- Sagem(Sag) 2016/2036 352×384 vantages, this paper concatenated the fingerprint fea- tures extracted by multiple CNNs. In addition, the un- Biometrika(Bio) 2000/2000 312×372 known feature of neural network extraction makes it im- Italdata(Ita) 2000/2000 640×480 LivDet2013 possible to find the optimization direction of the connec- Crossmatch(Cro) 2250/2250 800×750 tion feature. Regarding the issue above, we introduce a Swipe(Swi) 2200/2153 208×1500 FLD algorithm based on multi-modal features, and uses genetic algorithm to optimize these features. Through genetic algorithm, concatenated features are optimized adaptively and the difference features between real and 4.2. Performance evaluation ctiteria fake fingerprints are deeply mined. The experimental In the field of fingerprint liveness detection, the aver- results show that after the optimization of the genetic age classification error (ACE) is a widely accepted evalu- algorithm, the accuracy of FLD is improved by about 1% ation standard. The ACE is defined as the average value on the basis of the highest accuracy of the sub-network. of false reject rate (FRR) and false accept rate (FAR), cal- Compared with other FLD algorithms, our method also culated as Eq.(2). has better accuracy and stability. 𝐹 𝑅𝑅 + 𝐹 𝐴𝑅 𝐴𝐶𝐸 = 2 , (2) Acknowledgement Based on the original VGG16, AlexNet and In this paper, the two authors have equally im- ResNet18 models, the final fully connected layer are portant contributions. This work is supported by set to 512 dimensions.Through four different operations the National Key R& D Program of China under Table 2 The Average Classification Error of different models when datasets are LivDet2011 and LivDet 2013 repectively. Dataset Model The Average Classification Error ACE in (%) Bio Dig Ita Sag Average Ours 1.1 1.8 7.4 3.2 3.3 VGG16 [31] 2.1 2.6 8.4 4.4 4.3 AlexNet [31] 3 5.9 18.7 5.9 8.3 ResNet18 [33] 3.5 4.2 13.3 6 6.7 LivDet2011 CNN [28] 9.2 1.35 12.35 6.1 7.2 DRN [29] 9.6 1.9 13.5 6.4 7.8 MBLTP [32] 9.7 7.0 16.0 10.9 10.9 RWG [34] 5.7 6.2 9.4 3.14 6.11 LQF [35] 7.4 5.6 11.4 6.7 7.8 Bio Cro Ita Swi Ave Ours 2.8 12.6 0.7 4.8 5.2 VGG16 [31] 3.1 37.4 1.1 6 11.9 AlexNet [31] 5.6 15 1.3 5.9 6.9 ResNet18 [33] 5.1 10.8 4.5 6.9 6.8 LivDet2013 CNN [28] 4.35 7 1.4 5.1 4.4 DBN [30] 1.15 7.91 1.35 6.5 4.2 RWG [34] 0.6 — 0.85 3.2 1.55 LQF [35] 5.8 5.2 4.3 14.2 7.4 DCMBP [36] 0 10.62 0.85 15.65 6.78 grant 2018YFB1003205; by the Jiangsu Basic Research [6] Schuckers S A C., Spoofing and anti-spoofing mea- Programs-Natural Science Foundation under grant sures. Information Security Technical Report, vol. BK20200807; by the Research Startup Foundation of 7, no. 4, pp.56-62, 2002. NUIST 2020r015; by the Canada Research Chair Program [7] Nikam S B, Agarwal S, Texture and wavelet-based and the NSERC Discovery Grant; by the Priority Aca- spoof fingerprint detection for fingerprint biomet- demic Program Development of Jiangsu Higher Educa- ric systems Emerging Trends in Engineering and tion Institutions (PAPD) fund; by the Collaborative Inno- Technology, 2008. ICETET’08. First International vation Center of Atmospheric Environment and Equip- Conference on. IEEE, pp. 675-680, 2008: 675-680. ment Technology (CICAEET) fund, China. [8] Abhyankar A S, Schuckers S C, A wavelet-based approach to detecting liveness in fingerprint scan- ners Defense and Security. International Society References for Optics and Photonics, pp. 278-286, 2004. [9] Schuckers S, Abhyankar A, Detecting liveness in [1] Yambay D, Ghiani L, Denti P, Marcislis G L, fingerprint scanners using wavelets: Results of Roli F, and Schuckers S, Liv Det 2011-Fingerprint the test dataset Biometric Authentication. Springer liveness detection competition 2011. Iapr Interna- Berlin Heidelberg, pp. 100-110, 2004: 100-110. tional Conference on Biometrics. IEEE, April 2012, [10] Zhang Y, Tian J, Chen X, et al, Fake finger detection pp.208-215. based on thin-plate spline distortion model Ad- [2] Jia X, Yang X, Cao K, et al, Multi-scale local binary vances in Biometrics. Springer Berlin Heidelberg, pattern with filters for spoof fingerprint detection. pp. 742-749, 2007. Information Sciences: an International Journal, vol. [11] Moon Y S, Chen J S, Chan K C, et al. Wavelet based 268, pp. 91-102, 2014. fingerprint liveness detection. Electronics Letters, [3] Maltoni D, Maio D, Jain A K, et al, Handbook of 2005, 41(20): 1112-1113. fingerprint recognition. springer, 2009. [12] Antonelli A, Cappelli R, Maio D, et al, A New Ap- [4] Sousedik C, Busch C. Presentation attack detection proach to Fake Finger Detection Based on Skin methods for fingerprint recognition systems: a sur- Distortion. Advances in Biometrics, International vey. Biometrics Iet, vol. 3, no. 4, pp. 219-233, 2014. Conference, ICB, 2006. [5] Maltoni D, Maio D, Jain A K, et al.: Handbook of [13] Manivanan N, Memon S, Balachandran W, Auto- fingerprint recognition. Springer Science & Busi- matic detection of active sweat pores of fingerprint ness Media, 2009. using highpass and correlation filtering. Electron- [27] Nogueira R F, Lotufo R.D.A and Machado R C, ics letters, vol. 46, no. 18, pp. 1268-1269, 2010. ”Evaluating softwarebased fingerprint liveness de- [14] Nikam, Agarwal.: Local binary pattern and tection using Convolutional Networks and Lo- wavelet-based spoof fingerprint detection Interna- cal Binary Patterns,” Biometric Measurements and tional Journal of Biometrics, 2008, 1, pp. 141-159. Systems for Security and Medical Applications. [15] Kannala, Rahtu.: Bsif: Binarized statistical image IEEE, Rome, Italy, November, 2015, pp:22-29 features Proceedings of the 21st international con- [28] Yuan C , Xia Z , Sun X , et al. Deep Residual Net- ference on pattern recognition (ICPR2012). IEEE, work with Adaptive Learning Framework for Fin- 2012, pp. 1363-1366. gerprint Liveness Detection[J]. IEEE Transactions [16] Ghiani, Marcialis, Roli.: Fingerprint liveness de- on Cognitive and Developmental Systems, 2019, tection by local phase quantization, Proceedings PP(99). of the 21st International Conference on Pattern [29] Yuan, C., Xia, Z., Jiang, L., Cao, Y., Wu, Q. J., & Recognition (ICPR2012). IEEE, 2012, pp. 537-540. Sun, X. (2019). Fingerprint liveness detection using [17] Mohan, James.: Fingerprint Spoofing Detection us- an improved CNN with image scale equalization. ing HOG and Local Binary Pattern International IEEE Access, 7, 26953-26966. Journal of Advanced Research in Computer and [30] Kim, S., Park, B., Song, B. S., & Yang, S. (2016). Deep Communication Engineering, 2017, 6, pp. 586-593. belief network based statistical feature learning for [18] Nikam, Agarwal, Texture and wavelet-based spoof fingerprint liveness detection. Pattern Recognition fingerprint detection for fingerprint biometric sys- Letters, 77, 58-65 tems. Emerging Trends in Engineering and Tech- [31] Nogueira, Rodrigo Frassetto, Roberto de Alencar nology, 2008. ICETET’08. First International Con- Lotufo, and Rubens Campos Machado. Fingerprint ference on. IEEE, pp. 675-680, 2008: 675-680. liveness detection using convolutional neural net- [19] Yuan , Xia, Sun, et al, Fingerprint liveness detec- works. IEEE transactions on information forensics tion using multiscale difference co-occurrence ma- and security 11.6 (2016): 1206-1213 trix. Optical Engineering, vol. 55, no. 6, 2016. [32] X. Jia, X. Yang, Y. Zang, N. Zhang, R. Dai, J. Tian [20] Nogueira R F, Lotufo R.D.A., Machado R C, ”Finger- and J. Zhao, Multiscale block local ternary patterns print Liveness Detection Using Convolutional Neu- for fingerprint vitality detection. Conference: Bio- ral Networks,” IEEE Transactions on Information metrics (ICB), 2013 International Conference on, Forensics & Security, Vol.11, No.6, pp.1206-1213, June, 2013(ICB), pp:1-6 June, 2016. [33] Szegedy C , Ioffe S , Vanhoucke V , et al. Inception- [21] Yuan C , Xia Z , Sun X , et al. Deep Residual Net- v4, Inception-ResNet and the Impact of Residual work with Adaptive Learning Framework for Fin- Connections on Learning[J]. 2016. gerprint Liveness Detection[J]. IEEE Transactions [34] Xia, Zhihua, Lv, et al. Rotation-invariant Weber on Cognitive and Developmental Systems, 2019, pattern and Gabor feature for fingerprint live- PP(99). ness detection[J]. Multimedia Tools & Applica- [22] Zhang Y, Shi D, Zhan X,et al. Slim-ResCNN: A deep tions, 2018. Residual Convolutional Neural Network for Fin- [35] Sharma R P , Dey S . Fingerprint liveness detection gerprint Liveness Detection[J]. IEEE Access, 2019, using local quality features[J]. The Visual Com- PP(99):1-1. puter, 2018, 35(3). [23] Agarwal S, Chowdary C R, Sourabh V. EaZy Learn- [36] Yuan C , Sun X , Wu Q . Difference co-occurrence ing: An Adaptive Variant of Ensemble Learning for matrix using BP neural network for fingerprint Fingerprint Liveness Detection. 2021. liveness detection[J]. Soft Computing, 2018. [24] Yambay D, Ghiani L, Denti P, Marcialis G L, Roli F and Schuckers S,”LivDet 2011-Fingerprint live- ness detection competition 2011.” Iapr Interna- tional Conference on Biometrics. IEEE, New Delhi, India, April, 2012, pp.208-215. [25] Ghiani L, Yambay D, Mura V, Tocco S, Marcialis G L, Roli F and Schuckcrs S, ”Livdet 2013 finger- print liveness detection competition 2013,” Biomet- rics (ICB), 2013 International Conference on. IEEE, June, 2013, pp.1-6. [26] Samma H , Suandi S A . Transfer Learning of Pre- Trained CNN Models for Fingerprint Liveness De- tection[M]. 2020.