A comprehensive survey of fingerprint presentation attack detection

Anatomical or physiological features

This section describes state-of-the-art PAD methods, where anatomical or physiological features, such as sweat pores and perspiration, are utilized to determine if a fingerprint is real or artificial. The detection schemes are mainly focused on sweat pores [Figure 6] or perspiration. The term sweat pore describes tiny openings in the skin where sweat reaches the surface from their respective glands below. Perspiration is a phenomenon where sweat starts from the pores and scatters along ridges. Thus, regions between pores become darker. Modern PAD methods exploit this property and by observing numerous samples, they can capture a perspiration pattern that indicates whether a finger is real or not. Table 3 summarizes the presented methods.

Figure 6. (A) Fingerprint image and (B) pores^[56].

Tan and Schuckers^[21] proposed a detection scheme based on perspiration and utilized ridge signal and valley noise analysis. Gray level patterns in spatial, frequency and wavelet domains in combination with classification trees and neural networks were used for the detection. The proposed scheme achieved an EER of 0.9% on their own dataset comprised of 644 bona fide and 570 artificial fingerprint samples.

Espinoza and Champod^[61] proved that the pores of the skin of a fingerprint can be used as a feature that can discriminate bona fide and artificial fingerprints. The discriminative factor in their PAD scheme was the difference between the total number of pores in bona fide and artificial fingerprint samples. Their method achieved a 21.2% FAR and an 8.3% FRR using their own dataset.

Marasco and Sansone^[62] utilized a feature set combined of the residual noise of the fingerprint image to detect the coarseness of the artificial fingerprint, first-order statistics based on the gray level of each pixel, the intensity distribution to detect PAIs and the individual pore spacing, which is unique to every human. An SVM, decision tree, MLP and Bayesian classifier were chosen as classifiers, depending on the best performance per sensor. This method outperformed other approaches, exhibiting an ACE of 12.5% on the LivDet 2009 dataset, and offered significant gains in speed.

Pereira et al.^[63] used a combination of feature sets proposed in Refs.^[64,65]. The 17-dimensional feature vector was minimized with the sequential forward selection technique^[66]. A SVM and a MLP were used. The authors concluded that the proposed feature set improved performance by 33.56% and the ACE was increased as more attempts for authentication were allowed. The best performance was reached on a single attempt for the acceptance scenario and it exhibited an ACE of 4.17% for a SVM and 4.27% for a MLP classifier. Furthermore, the SVM performed better in general as a classifier, except on biometrics acquired from elderly people where the MLP performed better.

Marcialis et al.^[67] proposed the utilization of features that are encountered in the production of artificial fingers. They utilized the Fourier power spectrum of the fingerprint image and concluded that these features can be relevant for discriminating bona fide fingerprints from artificial ones. Moreover, the fusion of features extracted from PAIs and bona fide presentations showed a significant improvement in performance. Their method was evaluated on the LivDet 2009 dataset and showed promising results.

Johnson and Schuckers^[68] proposed a perspiration-based PAD. After detecting the pores, a small surrounding area of the pore was inspected to ascertain the perspiration activity. A SVM classifier with a radial basis function kernel was utilized and the method was evaluated on the LivDet 2011-2013 datasets. Experimental results showed that the method performed well, especially when combined with other PAD techniques. More specifically, the best approach exhibited an EER of 12% on the LivDet 2011 and an EER of 12.7% on the LivDet 2013.

Lu et al.^[69] proposed a method where after the extraction of pore information using a Mexican hat (Mexh) wavelet transform and adaptive Gaussian filters, five statistical pore distribution features were utilized. These features were the pore number (total amount of pores), pore density, mean pore space, variance and the variation coefficient. A SVM was used for classification. This technique exhibited an ACE of 7.11% on the LivDet 2011 and an ACE of 11.4% on the ATVS datasets. Table 3 summarizes the aforementioned PAD techniques.

Image quality features

In this section, we present the state-of-the-art PAD methods that seek to find detectable differences in the scans of a living finger in comparison to an artificial one. Measures like continuity, clarity and strength of valleys and ridges are being utilized for anti-spoofing. Typically, features extracted from ridge-valleys are utilized but other features were proposed as well, as shown in Table 4. The advantages of these methods are the simplicity, low computational complexity and fast response times^[14].

Tan^[70] proposed a PAD method that utilized noise analysis along the valleys in the ridge-valley structure of fingerprint images. Wavelet decomposition was utilized to acquire statistical features in multiresolution scales. Decision trees and neural networks were used for classification, while two datasets were used for evaluation. The first one was comprised of 58 live, 80 artificial and 25 cadaver samples, whilst the second included 28 bona fide and 28 artificial fingerprints. This method exhibited a correct classification rate from 90.9% to 100% depending on the technology of the scanner.

Galbally et al.^[71] proposed a PAD method that utilized ten different quality features of the image that depend on the ridge strength, continuity and clarity. LDA was used as a classifier. This method achieved an ACE of 6.56% on the LivDet 2009 dataset.

Lee et al.^[72] proposed a novel PAD method that measured the standard deviation of the fractional Fourier transform of a line, which was detected when a fingerprint image was transformed into the spatial frequency domain. This transformation was accomplished with the use of a two-dimensional fast Fourier transform. For a dataset of 3750 bona fide and artificial fingerprint samples in total, this method exhibited an error rate of 11.4% when a certain region was utilized after the fractional Fourier transform.

In Ref.^[73], a method based on the fusion of the spectral band, middle ridge and valley line was proposed. SVMs and quadratic classifiers were used for the classification. This system was tested on the Biometrics Engineering Research Center dataset^[74] and exhibited better security than other fingerprint recognition schemes. It also has the advantage that it uses only one fingerprint. A classification error of ~6% was exhibited with the utilization of all three proposed features.

In Ref.^[75], image quality features that depend on the ridge strength, directionality, continuity and clarity and the integrity of the ridge-valley structure or estimated verification were measured to classify a fingerprint as artificial or not. The classification was performed by a LDA classifier. More specifically, their method showed ACEs of 12.5% and 5.4% on the LivDet 2009 and ATVS datasets, respectively. Compared to other similar methods, the fact that only one sample is needed, makes the sample acquisition process faster and less invasive.

In another work of Galbally et al.^[76], 25 image quality features were utilized to discriminate bona fide samples from artificial ones. Their method was tested on the LivDet 2009 dataset and achieved an APCER of < 13% and a BPCER of ≤ 14%.

Sharma and Dey^[77] proposed an architecture that utilized five novel and eight existing quality features that depend on the ridge-valley shape and are sensor independent. Thus, a 13-dimensional feature vector was formed and a sequential forward floating selection (SFFS) and a random forest feature selection algorithm were deployed to select the optimal feature set. A SVM, random forest and gradient boosted tree were utilized for classification. This approach showed ACEs of 5.3% on LivDet 2009, 7.80% on LivDet 2011, 7.4% on LivDet 2013 and 4.2% on LivDet 2015.

Textural features

Researchers have found that the textural features [Figure 7] of the fingertip, such as smoothness and morphology, can be used to distinguish real fingerprints from artificial ones^[78,79]. Methods belonging to this category make use of such textural features. The extracted features are then presented as input to a classifier, which in most cases is an SVM. These methods are described below and summarized in Table 5.

Figure 7. Textural characteristics of a real fingerprint image^[22].

Nikam and Agarwal^[22] utilized curvelet features, such as energy, co-occurrence and fused signatures, to discriminate bona fide samples from artificial ones. To limit the dimensionality of the feature vector, they applied an SFFS algorithm. An ensemble classifier, on the basis of the “majority voting rule”, of three independent classifiers, namely, AdaBoost.M1, an SVM and an alternating decision tree, was used for classification. Their method was tested on their own dataset comprising of 185 bona fide and 240 artificial samples and on the FVC 2004^[80] dataset comprised of only bona fide samples. When the fusion of energy and co-occurrence signatures occurred, the proposed technique achieved a 99.29% correct classification rate, which outperformed the wavelet and power spectrum PAD techniques.

Ghiani et al.^[81] proposed a new feature set, extracted by the deployment of the textural analysis of the acquired image spectrum, known as rotation invariant local phase quantization (LPQ). This method exhibited an average EER of 12.3% on the LivDet 2011 dataset.

Gragnaniello et al.^[82] used textural classification by utilizing the Weber local descriptor (WLD). A linear kernel SVM classifier was used for classification. The classifier was trained on discriminative features build from the joint histograms of the differential excitation and orientation of every pixel of the acquired sample. This method presented good performance and it was further improved with the integration of other LPQ descriptors, especially if the latter relied on different image attributes. The ACE concerning WLD was 2.95%, while when the WLD and LPQ were combined, the ACE was 1.14% on the LivDet 2009 dataset. Experimental results on the LivDet 2011 dataset showed an ACE of 15.33% for the WLD, while the ACE for the fusion of the WLD and LPQ descriptors was 7.86%.

Jia et al.^[83] developed a descriptor known as multi-scale block local ternary patterns that depend on the average value of the pixel blocks. The differences amongst the pixels and the threshold constitute the ternary pattern. This method showed an ACE of 9.8% on the LivDet 2011 dataset.

Pereira et al.^[84] proposed spatial surface coarseness analysis (SSCA). SSCA is based on wavelet analysis of the fingerprint surface with the addition of spatial features. A polynomial kernel SVM was used for classification. SSCA exhibited an ACE of 12.8% on the LivDet 2011 dataset.

Ghiani et al.^[85] introduced a novel descriptor known as binarized statistical image features (BSIFs). This descriptor encodes the local fingerprint texture on a feature vector. This method was evaluated on the LivDet 2011 dataset and achieved an EER of 7.215% when a 7 × 7 window size was used in conjunction with a 4096-dimensional feature vector.

Jia et al.^[86] introduced a novel descriptor known as a multi-scale local binary pattern (MSLBP). The MSLBP can be implemented in two ways and both can achieve good performance for PAD. They tested MSLBP on the LivDet datasets and the best MSLBP implementation exhibited an ACE of 7.475%.

Gragnaniello et al.^[87] developed a method based on the wavelet-Markov local descriptor. The feature vector was formed by exploiting joint dependencies among wavelet coefficients. To limit the dimensionality of the feature vector, principal component analysis (PCA) was used. An SVM classifier achieved an ACE of 2.8% on the LivDet 2009 dataset.

Zhang et al.^[88] proposed a method that depends on wavelet analysis and LBP. Wavelet analysis was applied to produce the denoised image and the residual noise image. The LBP histograms were constructed based on the residual noise and denoised images. An SVM based on a polynomial kernel was deployed for classification. This approach offered ACEs of 11.47% on the LivDet 2011 and 11.02% on the LivDet 2013.

Gottschlich et al.^[89] proposed a descriptor known as histograms of invariant gradients. Fingerprint discrimination was based on multiple histograms of invariant gradients, computed from spatial neighborhoods within the fingerprint. The best variation of the proposed method achieved an ACE of 12.2% on the LivDet 2013 dataset.

Jiang and Liu^[90] used co-occurrence matrices from image gradients for feature extraction. The image was first quantized to decrease the dimensionality and increase the usefulness of the feature vector. Afterwards, the image differences were calculated from adjacent quantized pixels along the horizontal and vertical axes. These differences were then truncated in a range within a specific threshold. Finally, the truncated differences were used to produce the co-occurrence matrices, which were utilized as features. An SVM was trained on two datasets and the proposed method achieved ACEs of 6.8% and 10.98% on the LivDet 2009 and 2011 datasets, respectively.

Gragnaniello et al.^[91] developed a new local descriptor known as the local contrast phase. After the analysis of the acquired sample in the spatial and frequency domains, information on the local amplitude contrast and local behavior of the image was gathered depending on the selected transform coefficients. The two-dimensional contrast-phase histogram crafted from the information collected in the previous stage was used as a feature vector. A linear-kernel SVM classifier was utilized and an ACE of 5.7% was found for the LivDet 2011 dataset.

Gottschlich^[92] introduced a novel local image descriptor known as the convolution comparison pattern. This descriptor utilized rotation invariant image patches to compute the discrete cosine transform (DCT) and the comparison of pairs of two DCT coefficients to obtain binary patterns, which are summarized into histograms, comprised of the relative frequencies of pattern occurrences. The feature vector was acquired by the concatenation of multiple histograms and the classification was performed by an SVM. This descriptor, with the use of a specific configuration, achieved an accuracy of 93% on the LivDet 2013 dataset.

Dubey et al.^[93] used low level gradient features collected with the utilization of speeded-up robust features and the pyramid extension of the histograms of oriented gradient in conjunction with textural features acquired with the use of Gabor wavelets. This architecture exhibited an EER of 3.95% on the LivDet 2011 and achieved an ACE of 2.27% on the LivDet 2013 dataset.

Yuan et al.^[94] proposed the angular second moment, entropy and inverse differential moment and correlation to form the feature vector. These parameters were used as textural features and were extracted from eight difference co-occurrence matrices. The proposed method achieved ACEs of 15.32% on the LivDet 2013 dataset, 9.28% on the LivDet 2011 dataset and 7.92% on the LivDet 2009 dataset.

Kim and Jung^[95] proposed the local accumulate smoothing pattern descriptor. The feature vector of this classifier was based on the local textural patterns, as they were represented by the accumulated smoothing space. A linear SVM classifier was used for classification. This architecture achieved a classification rate of 88.49% on the LivDet 2009 dataset and of 78.78% on the LivDet 2011 dataset.

Ghiani et al.^[96] suggested a descriptor entitled BSIFs. BSIFs share similarities to LBPs and LPQ, but the formed feature vector is based on a set of filters that are learnt from natural images. The proposed descriptor achieved ACEs of 3.03% on the LivDet 2009 dataset, 9.62% on the LivDet 2011 dataset and 5.71% on the LivDet 2013 dataset.

Kim^[97] defined and utilized Local coherence patterns (LCPs) [Figure 8] as features in the dominant direction of the captured fingerprint image to determine if a fingerprint was artificial or not. This method exploited the fact that artificial fingerprints tend to disperse differently in the image gradient field than the bona fide ones. The classification was performed by a linear SVM. The LCP descriptor exhibited an accuracy of 93.49% on the ATVS Dataset and an accuracy of 78.02% on the LivDet 2009, 2011, 2013 and 2015 datasets.

Figure 8. Procedure for LCP. (A) Input; (B) filtered image; (C) LCP image; (D) feature extraction^[97]. LCP: Local coherence pattern.

Kumpituck et al.^[98] proposed the wavelet-based local binary pattern, which is based on the utilization of the LBP for capturing the local appearance of the sub-band images. Prior to the utilization of the LBP, the fingerprint image was decomposed by the two-dimensional discrete wavelet transform. An SVM was used for classification. The proposed method achieved an ACE of 9.95% on the LivDet 2009-2013 datasets.

Xia et al.^[99] proposed a method that utilized co-occurrence matrices constructed from image gradients. Second and third-order co-occurrence matrices were used to train an SVM classifier. When third-order co-occurrence matrices were utilized as features, the proposed architecture achieved ACEs of 6.2% on the LivDet 2009 and 6.635% on the LivDet 2011.

González-Soler et al.^[100] proposed a method based on the bag of words algorithm. After extracting the scale invariant feature transform (SIFT) features, they encoded them with a spatial histogram of visual words. The classification was performed by a SVM through a feature map. This architecture achieved an ACE of 4.7% on the LivDet 2011 dataset.

Kundargi and Karandikar^[101] proposed a LBP texture descriptor with a wavelet transform that utilized the textural characteristics that differ in bona fide and artificial samples due to variations at the gray level of the image. The fingerprints were classified by linear and RBF kernel SVM classifiers. This method offered an ACE of 8.3% on the LivDet 2011 dataset.

Jiang and Liu^[102] introduced a method that utilized the uniform local binary pattern in three-layer spatial Gaussian pyramids. This method achieved a 21.205% ACE on the LivDet 2011 dataset.

Mehboob et al.^[103] proposed a novel descriptor known as the combined Shepard magnitude and orientation. This method extracts the global features of the fingerprint by computing the relation between the perceived Shepard magnitude and initial pixel intensities in the spatial domain. To achieve this, the fingerprint is considered as a two-dimensional vector. The descriptor first constructs perceived spatial stimuli by combining the logarithmic function of initial pixel intensities and the Shepard magnitude (SM) in the spatial domain. Next, the phase information (CO) is computed in the frequency domain. Finally, the SM and CO are concatenated and represented as a two-dimensional histogram. The rotation invariant version of LPQ was utilized for characteristic orientation computation. An SVM was used for classification and average error rates of 5.8%, 2.2% and 5.3% on the LivDet 2011, 2013 and 2015 datasets were achieved, respectively.

Xia et al.^[104] also suggested a novel local descriptor entitled the Weber local binary descriptor (WLBD) that consists of two components that were used to extract intensity-variance and orientation features. The first is the local binary differential excitation module that captures the spatial structure of the local image patch and the second is the local binary gradient orientation module that is designed to extract gradient orientation from center-symmetric pixel pairs. The output of these components formed a discriminative feature vector [Figure 9] that was used as the input for an SVM classifier. The WLBD achieved ACEs of 9.67% on the LivDet 2015 dataset, 1.89% on the LivDet 2013 dataset and 5.96% on the LivDet 2011 dataset.

Figure 9. (A) Live fingerprint image. (B) WLD-DE, (C) LBP and (D) LBDE calculated from (A)^[104].

In Ref.^[105], a method utilizing the guided filtering and hybrid image analysis was proposed. After performing ROI extraction and guided filtering to acquire the denoised image, the co-occurrence of adjacent LBPs (CoALBPs)^[106] descriptor was utilized to form the feature vector from the original and the denoised image. An SVM with an RBF kernel was used and the proposed method exhibited average accuracies of 94.33% on the LivDet 2011 and 98.08% on the LivDet 2013 datasets. Moreover, the computation time was 4.5 times faster in comparison to deep learning methods.

Kumar and Singh^[107] proposed a fingerprint authentication system with a PAD module, utilizing supervised learning with minutiae extraction and classification with an SVM. Features like homogeneity, contrast, energy, entropy and mean last histogram were extracted. The proposed module achieved an average performance of 96.06% on the FVS^[108] and ATVS datasets.

Neural networks

Neural networks have been used for PAD with great success. Most of these methods share the similarity of the segmentation of the background or foreground of the fingerprint image and the extraction of local patches of the image that include the ROI. In recent years, researchers utilized deep learning methods to detect PAs. Deep neural networks like CNNs have been extensively utilized for several security tasks like steganalysis^[109,110], but nowadays they are also used for fingerprint PAD. CNNs are either utilized as feature extractors or even perform the classification. There are also proposed methods in the literature that use transfer learning. Transfer learning is the reuse of a pre-trained neural network on a new problem, i.e., it exploits the knowledge gained from a previous task to improve the generalization to another. Other methods exploit either generative adversarial networks (GANs) or restricted Boltzmann machines for PAD. Table 6 summarizes the state-of-the art methods that utilize a neural network (shallow or deep) as a feature extractor or classifier.

Menotti et al.^[111] proposed a detection technique that utilized neural networks. One of their approaches utilized an optimized convolutional neural network and an SVM for classification. A second approach was based on the backpropagation algorithm for filter optimization. They concluded that the combination of the two approaches performed better and achieved an ACC of 98.97% on the LivDet 2013.

Nogueira et al.^[112] proved that pretrained CNNs achieve high accuracy in PAD. In their work, they analyzed the false fingerprint detection performance of VGG^[39] and AlexNet^[113] [Figure 10] that were trained on natural images and further tuned with fingerprint samples. The LivDet 2009, 2011 and 2013 datasets were used and the proposed CNN-VGG showed an ACE of 2.9% and won the LivDet 2015 competition.

Figure 10. AlexNet^[113].

Kim et al.^[114] proposed an architecture based on a deep belief network (DBN) with multiple layers of a restricted Boltzmann machine, trained on a set of bona fide and artificial samples. The proposed DBN, when trained with augmented data, achieved an average accuracy of 97.10% on the LivDet 2013 dataset.

Marasco et al.^[115] experimented on the effectiveness of CNNs as PAD methods. They tested three CNNs that achieved the best accuracy on the LivDet 2013 dataset, namely, CaffeNet (96.5%), GoogLeNet (96.6%) and Siamese (93.1%). They have also indicated that CNNs exhibited the ability to successfully adapt to PAD problems (if they were pre-trained on the ImageNet dataset^[116]) and achieved an area under the curve in the range of -3.6% to +4.6%.

Pala and Bhanu^[117] proposed a method based on a triplet of CNNs. Their method employs a variant of the triplet objective function that considers representations of fingerprint images, where the distance between feature points is used to discriminate artificial and bona fide samples. Their architecture scored an ACE of 1.75% on the LivDet 2009, 2011 and 2013 datasets.

Chugh et al.^[18] proposed a method that was based on CNNs and minutiae information. For the training of the CNN, aligned and centered local patches (96 × 96 pixels) were utilized. They defined a “spoofness” score, which is the output of the softmax layer of the trained CNN. The “spoofness” score had a range of 0 to 1, where 1 denoted that the sample was artificial. They also crafted the Fingerprint Spoof Buster, which is a graphical user interface that permits the visual evaluation of the fingerprint by the human operator of the scanner. The proposed CNN was evaluated on the LivDet 2011, 2013 and 2015 datasets, as well as the MSU-FPAD and Precise Biometrics Spoof-Kit datasets. Experimental results showed a significant reduction in error rates, with an APCER lower than 7.3% and a BPCER of 1%.

Jung and Heo^[118] proposed a method, where the proposed CNN was trained directly from fingerprints, used the squared error layer and had no fully connected layer. The proposed architecture presented a higher average accuracy of 95.3% on the LivDet 2015 dataset compared to other existing methods.

Pinto et al.^[119] utilized deep learning and concluded that it achieves very good performance in PAD tasks. The main drawback was the poor performance of deep learning when encountered PAs with instruments that were either not included at all or not included with a satisfactory number of samples in the training data. They also concluded that the best reliability is presented when models are based on both “hand-crafted” and “data-driven” solutions.

Park et al.^[120] proposed a patch-based PAD method that utilized a fully convolutional neural network, the so-called Fire module of SqueezeNet, which had fewer parameters and demands, with only 2.0 Mb of memory required. This method utilized an optimal threshold, as opposed to the voting method, which decreased the misdetection rate. This architecture achieved an ACE of 1.35% on the LivDet 2011, 2013 and 2015 datasets, when the training was carried out through data augmentation and the patch size was 48 × 48 pixels.

Park et al.^[121] proposed a convolutional network known as the patch-based CNN because it utilizes patches (small ROIs) of the fingerprint image. They used three categories for classification, i.e., live, artificial and background. The proposed architecture, when used on 48 × 48 pixel patches, achieved an ACE of 1.43% on the LivDet 2011, 2013 and 2015 datasets in a radically reduced execution time.

Zhang et al.^[122] proposed a lightweight residual convolutional neural network (Slim-ResCNN) that requires less processing time and is robust to overfitting. Local patch extraction was based on statistical histograms and the center of gravity. The proposed method exhibited an overall accuracy of 95.25% on the LivDet 2017 dataset.

Yuan et al.^[123] proposed a real-time fingerprint PAD method based on autoencoders. Automatic feature extraction was performed with the use of a stacked autoencoder. Unsupervised learning was used for pre-training, while the detection was performed with the use of supervised training. A Softmax classifier was utilized for classification. This method achieved ACEs of 19.62% on the LivDet 2013 and 18% on the LivDet 2015.

Pereira et al.^[124] tackled the PAD generalization problem, especially for PAs performed with materials not seen in training, with the use of an adversarial training methodology. The proposed method was evaluated with a MLP classifier and a CNN. The regularized CNN approach achieved an APCER of 0.60% on the LivDet 2015 dataset.

Uliyan et al.^[125] utilized discriminative restricted Boltzmann machines (DRBMs) in combination with a deep Boltzmann machine (DBM) to extract deep features from the acquired samples. A K-NN classifier was used for classification. The proposed DRBM-DBM architecture exhibited an ACE of 3.6% on the LivDet 2013 dataset.

Zhang et al.^[126] proposed a lightweight CNN known as FLD to achieve improved PAD on new materials and to minimize complexity. To address the issue of global average pooling, an attention pooling layer was used. Moreover, a novel block structure (Block D&R) was introduced where the residual path is integrated into the original dense block. FLDNet exhibited ACEs of 1.76%, over all sensors, on the LivDet 2015 dataset and 0.25% on the LivDet 2013 dataset.

Jian et al.^[127] proposed a densely connected convolutional network (DenseNet) along with a genetic algorithm utilized for network optimization. The genetic algorithm can automatically optimize DenseNet by finding the optimal structure from the solution space. The proposed model achieved 98.22% accuracy on a testing set containing the LivDet 2009-2015 datasets.

Fusion of features

In this category, approaches that combine features from the different aforementioned categories are presented. Moreover, hybrid methods that utilize both static and dynamic features will be discussed^[128-134]. These methods exploit the advantages and minimize the drawbacks of the aforementioned PAD methods, as these were presented in Dynamic methods, Anatomical or physiological features, Image quality features, Textural features and Neural networks. A summary of the presented methods is given in Table 7.

Derakhshani et al.^[128] proposed a method that detected the perspiration phenomenon and it was based on static and dynamic features acquired from two fingerprint samples captured with a time interval of 5 s. One static and four dynamic features were used as input to a back-propagation neural network, which was utilized for classification. The dataset for training and testing was comprised of 18 sets of fingerprint samples from live individuals, 18 from cadavers and 18 from spoof materials. On this dataset, the proposed scheme achieved an accuracy of 100%.

Parthasaradhi et al.^[129] proposed a method that depends on static features and on the changes due to perspiration to fingerprint images taken at 0, 2 and 5 s. By deploying a weight decade method during the training of a neural network classifier, they concluded that there was significant improvement in performance. On an image sequence dataset of 33 live subjects, 14 cadaver fingers and 33 artificial samples, their method achieved an FLR of 0% and an FSA in the range of 0%-18.2% depending on the sensor technology.

In an extension of their previous work^[129], Parthasaradhi et al.^[130] used several classification methods, a shorter time window, a more diverse dataset and included other fingerprint sensor technologies. One static and six dynamic measures were used for classification and all classifiers achieved ~90% accuracy. More specifically, on a dataset of 75 samples per scanner, the proposed method achieved an FLR in the range of 6.77%-20% and an FSA in the range of 5%-20% for an optical scanner, and an FLR of 0%-26.9% and an FSA of 4.6%-14.3% for a capacitive scanner. Finally, the same method exhibited an FLR of 6.9%-38.5% and an FSA of 0%-19% for electro-optical scanners.

Tan and Schuckers^[131] proposed a PAD method that depends on the static and dynamic features acquired from intensity histograms of the 0 and 5 s images of a fingerprint. On a dataset of sequences of images (30 bona fide, 40 artificial and 14 cadaver fingerprints) captured by three different scanners, their method exhibited an FLR of 0% and an FSA of 8.3% for an optical sensor. The same method exhibited an FLR of 6.7% and an FSA of 0% for a capacitive sensor, and an FLR of 7.7% and an FSA of 5.3% for an electro-optical sensor.

Tan and Schuckers^[132] extended their previous work^[131], by reducing the time between capturing fingerprint samples to 2 s instead of 5 s. On a dataset of 58 live, 50 artificial and 25 cadaver fingerprints, the augmented method showed an accuracy of 90%-100% for some scanners by using a classification tree.

Jia and Cai^[133] proposed an extension of their previous work^[50]. In this study, five features were utilized. Two of them represented skin elasticity, whilst the other three represented perspiration. This method computed two static features, while the remaining three were dynamic features. The proposed scheme was tested on a dataset comprising of 770 image sequences and achieved an EER of 4.49%.

Plesh et al.^[134] used a sensor with time-series and color sensing capabilities to capture a gray scale static image and a time-series color capture simultaneously. A dynamic color capture has the ability to measure signs, such as the perspiration, skin elasticity deformation and the displacement of blood that occurs when a finger is pressed. In their work, the second capture (the dynamic one) was utilized for classification with two methods. Initially, static-temporal feature engineering was utilized and then the InceptionV3 CNN^[135] trained on ImageNet was used for classification. In their work, the classification performance was evaluated with the use of a fully connected DNN utilizing solely static or dynamic features and a fusion of the two feature sets. On a custom dataset comprising of over 36,000 image sequences and a state-of-the-art set of PA techniques, the approach that utilized the fusion of both static and dynamic features achieved a mean APCER, at a 1.0% BPCER operation point, of 3.55 %, a mean APCER, at a 0.2% BPCER operation point, of 0.626% and a standard deviation, at 1.0% BPCER, of 1.96%.

Nogueira et al.^[136] evaluated the efficiency of two feature extraction techniques with data augmentation on an SVM classifier. After the preprocessing step, two feature extraction techniques, i.e., a CNN and LBP, were conducted and randomized PCA was utilized for dimensionality reduction. The proposed CNN, with the use of augmented data, exhibited an ACE of 4.71% on the LivDet 2009, 2011 and 2013 datasets.

Yuan et al.^[137] suggested a backpropagation neural network that utilized gradient values calculated by the Laplacian operator. They also presented a system that used the methods shown in Ref.^[94] and Ref.^[138] to create more productive input data and category labels. This architecture offered an ACE of 6.78% on the LivDet 2013 dataset.

Agarwal and Chowdary^[139] proposed the utilization of stacking and bagging ensemble learning approaches in fingerprint PAD. The suggested algorithms considered the similarities of datasets utilized for PAD. Moreover, they are adaptive because they comply with the features extracted from bona fide and artificial fingerprint samples. The proposed algorithms achieved better performance in accuracy and false positive rate than the best individual base classifier. More specifically, the stacking average accuracy was 80.76%, while the bagging average accuracy was 75.12% on the LivDet 2011 dataset.

Anusha et al.^[140] proposed an architecture that utilized global image and local patch features. It used LBP and Gabor filters in the preprocessing process to extract features using DenseNet^[141]. For the extraction of local patch features, a second DenseNet was used in combination of a channel and spatial attention network module^[142]. For patch discrimination, a novel patch attention network was proposed. This network was also used for feature fusion. This method showed average accuracies of 99.52%, 99.16% and 99.72% on the LivDet 2017, 2015 and 2011, respectively.

Agarwal and Bansal^[143] proposed the fusion of pores, perspiration and textural features for PAD. The dimensionality of the extracted feature vector was reduced with the use of a stacked autoencoder pretrained in a greedy layer wise manner. A supervised trained Softmax classifier was utilized for classification. In terms of performance, this method achieved ACEs of 0.1866% on the LivDet 2013 and 0.3233% on the LivDet 2015 dataset.

Li et al.^[144] used features extracted with three algorithms, i.e., SIFT, LBP and HOG. As a result, the benefits of these algorithms were combined and the overall performance was increased. A fusion rule was developed to fuse the features and the produced feature vector was then classified by an SVM. This method showed ACEs of 4.6% on the LivDet 2011 dataset, 3.48% on the LivDet 2013 dataset and 4.03% on the LivDet 2015 dataset.

Sharma and Selwal^[145] proposed a method that utilized majority ensemble voting based on three local and adaptive textural image features. The features were acquired with a new LBP variant descriptor called the local adaptive binary pattern (LABP), combined with features gathered with the use of a CLBP and BSIF descriptors. This method achieved ACERs of 4.11%, 3.19%, 2.88% and 2.97% on the LivDet 2009, 2011, 2013 and 2015 datasets, respectively.

Generalization efficient/wrapper methods

In this category, PAD methods are presented that focus on the efficiency against PAIs made with materials not used during training. Moreover, the performance of methods that were evaluated on novel PAI materials are reported. Some of the presented PAD methods can be used as add-ons or wrappers to any PAD method, in order to improve the performance against PAIs of unknown materials. A summary of the presented methods is given in Table 8.

Rattani and Ross^[146] proposed the creation of a novel material detector that detects PAIs made of novel materials. These samples were then used to automatically retrain and update the PAD method. To keep the computational complexity low, the automatic adaptation procedure was executed when the presentation attack detector was offline. This scheme was evaluated on the LivDet 2011 dataset and it exhibited an average correct detection rate of up to 74% and an up to 46% improvement in presentation attack performance when the adaptive approach was utilized.

Jia et al.^[147], in order to address the issue of lack of knowledge of the materials used for artificial fingerprints, suggested a one-class SVM with negative examples (OCSNE). The OCSNE showed an ACE of 23.6% on a modified dataset according to the needs of the evaluation the LivDet 2011, which was better than an SVM.

Rattani et al.^[148] proposed to handle PAD as an open set recognition problem. The authors claimed that their approach is useful, because during deployment novel materials different than the ones that the system was trained on may be used to construct artificial fingerprints. In their work a Weibull-calibrated SVM (W-SVM) was used as a novel material detector and as a PAD. They also developed a scheme to automatically retrain and update the W-SVM depending on the artificial fingerprints it detects. This approach was tested on the LivDet 2011 dataset and it exhibits an up to 44% improvement in performance.

Sequeira and Cardoso^[149] evaluated several classification methods and concluded that semi-supervised classification based on a mixture of Gaussians models yields better results. Moreover, they proposed the isolation of the fingerprint from the background by adding an automatic segmentation stage to the detection algorithms. The best method they evaluated exhibited an ACE of 8.35% on the LivDet 2013 datasets.

Nogueira et al.^[112] tested their PAD method against attack with PAIs not seen in training and their method based on a CNN-VGG achieved average ACEs of 16.1% on the LivDet 2011 dataset and 5.45% on the LivDet 2013 dataset.

Ding and Ross^[150], in their work based on performance metrics on the LivDet 2011 dataset, proved that using an ensemble of one-class SVMs based on descriptors that utilize different features achieves better accuracy than binary SVMs and also competed in performance with other state-of-the-art PAD algorithms immune to fabrication materials. Another advantage of this method is the limited number of artificial fingerprints used for training. The proposed method achieved an average correct detection rate on known PAI of 86.1%, while it presents an average correct detection rate on unknown materials of 84.7%, which is higher than the automatic adaption method presented in^[148].

Pala and Bhanu^[117] also evaluated their PAD scheme against unknown attacks and their method achieved average ACEs of 10.05% on the LivDet 2011 and 3.35% on the LivDet 2013.

Gajawada et al.^[151], to improve the efficiency of any PAD, especially against new materials, developed the Universal Material Translator (UMT) as a deep learning augmentation wrapper. They proposed the synthetization of artificial samples by utilizing only a small part of them. Along with the UMT, they also used a GAN. Although the authors believe that the combination of a UMT and a GAN produces better results, GANs insert certain artefacts and noise in the generated images that are detectable and negatively affect the performance of the classifiers. Their method was tested on the LivDet 2015 dataset and demonstrated a BPCER1000 of 21.96% on unknown PAIs.

Chugh and Jain^[152] experimented on the efficiency of the so-called spoof buster, a PAD wrapper developed in^[18], which could be used on top of any PAD method to improve generalization and efficiency, especially against PAIs not seen in training. The spoof buster used a deep convolutional neural network that utilized local patches centered and aligned using fingerprint minutiae. The MSU-FPAD dataset was utilized and their method achieved a (weighted average generalization performance) TDR of 75.24% when the leave-one-out method was used, as opposed to a TDR of 97.20% with an FDR of 0.2% when all PAIs were used in the training.

Engelsma and Jain^[153] proposed the utilization of three GANs on a training set that contained only bona fide samples. Their method showed improvement in cross-material performance, compared to one class or binary state-of-the-art classifiers, on their own dataset that contains 5531 artificial samples (from 12 materials) and 11,880 bona fide samples. For this dataset, it achieved a BPCER500 of 50.20%.

A Slim-Res CNN^[122] also demonstrated, to some extent, robustness against attack with new materials and presented an accuracy of 96.82% on the LivDet 2015 dataset, whose testing sets consisted of artificial fingerprints made of unknown materials.

Park et al.^[121] also evaluated their proposed method on the LivDet 2015 and concluded that their method presents efficiency against attacks made from unknown materials. The patched based CNN exhibited an ACE of 1.9%.

Grosz et al.^[154] suggested the use of adversarial representation learning in DNNs. Their proposal can be added to any CNN that uses 96 × 96 aligned minutiae-centered patches for training, along with the utilization of a style transfer network wrapper. Their method achieved a 92.94% TDR and a 0.2% FDR on the LivDet 2011 and 2015 datasets and on the MSU-FPAD dataset.

Zhang et al.^[126] also tested their proposed method against attacks from new materials and achieved an ACE of 3.31% on the LivDet 2015.

González-Soler et al.^[155] suggested three techniques for PAD. Initially, the pyramid histogram of visual words was utilized for extracting the local features of the fingerprint with the use of dense SIFT descriptors. Afterwards, the feature vector was formed with the utilization of three methods: (1) bag-of-words; (2) Fisher vector (FV); and (3) vector locally aggregated descriptors. A linear SVM was used for classification. The FV encoding achieved the best detection accuracy on the LivDet 2019 competition. Fusion of the three encodings achieved even better performance and yielded a BPCER100 in the range of 1.98%-17% in the presence of unknown PAI species.

Chugh and Jain^[156] proposed the utilization of the Universal Material Generator (UMG) for the performance improvement of any PAD method against unknown materials [Figure 11]. The UMG is a CNN, trained on characteristics of known materials that artificial fingerprints are made off, in an effort to synthesize artificial samples of unknown materials. The UMG improves the performance by increasing the TDR from 75.24% to 91.78% when the FDR was 0.2% and the average cross-sensor presentation attack detection performance from 67.60% to 80.63% on the LivDet 2017 dataset.

Figure 11. Artefacts and their respective images^[156].