Deep fingerprint classification network

Fingerprint is one of the most well-known biometrics that has been used for personal recognition. However, faked fingerprints have become the major enemy where they threat the security of this biometric. This paper proposes an efficient deep fingerprint classification network (DFCN) model to achieve accurate performances of classifying between real and fake fingerprints. This model has extensively evaluated or examined parameters. Total of 512 images from the ATVS-FFp_DB dataset are employed. The proposed DFCN achieved high classification performance of 99.22%, where fingerprint images are successfully classified into their two categories. Moreover, comparisons with state-of-art approaches are provided. This is an open access article under the CC BY-SA license.

In the biometric recognition system, there is no need to carry a key or card and no need to remember a PIN code or password. These facilitate the validation processes of users [1,24]. However, several drawbacks have been presented in biometric systems which include but not limited to their lower secrecy and security that make them vulnerable to external threats [1,25,26]. Fake biometrics are the most popular threats against biometric security systems. For instance, fake fingerprints can easily be constructed by using gummy fingers made from specific materials like silicone, wood glue and gelatine [27]. Figure 1 shows samples of real and fake fingerprints from ATVS-FakeFingerprint DataBase (ATVS-FFp_DB) [1] which are used in this study.
These challenges have increased the demand of developing robust biometric recognition systems [28,29]. Basically, fingerprint classification techniques of spoofing detection can be categorized into two classes based on additional sensors that are used or not:  These are the fingerprints readers and classifiers which are able to scan the  fingers with high resolution images for detecting the spoof and real fingerprints [30,31]. The aim of using the hardware classification techniques is to detect the main liveness features such as the skin distortion [32] and blood flow [33]. (b) Software based techniques: These methods can classify the fingerprints into fake and real fingerprints based on their scanned images by the used sensors without the need to utilize more hardware. This classification can also be divided into two categories: − Feature extraction and recognition methods: they depend on liveness detection of fingerprints.
However, due to their low accuracies for several spoofing materials and also the required long processing time for extracting features, many researchers have become less interest in this type [23,31].

−
Deep learning methods: these on the other hand have made great success in the field of recognition and classification such as fingerprints [34], emotions [35,36], bone fissures [37] and disorder classifications [38]. Deep learning technique is based on using adequate number of training data which lead to automatically learn their structures and features [34,39]. The classification processes in this field are most likely concentrate on either classifying the fingerprints into their features (right loop, left loop, arch and whorl) or classifying the fingerprints into real and fake. There are various approaches that were suggested for fingerprint detections. Examples of these are: an approach of using statistical weight calculation with non-reference fingerprint was presented in [24], an alternative approach for generating cancellable fingerprint was described by considering operations in matrices [15] and fingerprints can be used for advanced encryption to generate fuzzy vault crypto biometric key [16]. This paper proposes an efficient model of deep learning termed the DFCN to achieve an accurate classification of fingerprints into real and fake. The remaining sections in this paper are organized as follows: section 2 presents the prior work of fingerprint classifications, section 3 displays the methodology of the proposed model, section 4 clarifies the employed dataset specifications, section 5 demonstrates the results and discussions and section 6 yields the conclusion.

PREVIOUS WORK
In the literature, deep learning techniques that were used to classify fingerprints into real and fake can recently be found. Previous studies can be reviewed as follows: In 2016, Nogueira et al. used fine-tuned VGG-S and VGG-F convolutional neural networks (CNNs) to classify fingerprints into four classes (right loop, left loop, arch and whorl) using the NIST special database 4 (NIST SD4). The achieved classifying accuracies for both networks (the VGG-S and VGG-F) were 95.05% and 94.4%, respectively [27]. In 2018, El Hamdi et al. also used the NIST SD4 database to test their proposed model. The model was constructed based on the combination of conic radon transform (CRT) and CNN. The results showed high classification accuracy of 96.5% [40]. A patch-based model using a fully CNN with optimal threshold to detect spoof fingerprint was utilized by Park et al. in 2018. They employed the liveness detection (LivDet) datasets (LivDet-2011(LivDet- , 2013(LivDet- , 2015 in the experiments. Their proposed method revealed high performance with 1.35% average classification error [41].

895
(DBM) network and used the LivDet-2013 dataset. The proposed model was mainly consisted of three steps: starting with image normalization followed by training process and ended with the classifying process of support vector machine (SVM). The accuracy of the suggested system was 85.82% [42]. Also in 2019, Yuan et al. achieved satisfactory classifying results by using a deep model of deep residual network (DRN) followed by an enhanced and classified method named the local gradient pattern (LGP) [43]. Last but not least, Zhang and his researching team proposed a simple deep system depending on residual CNN called the Slim-ResCNN to classify the fingerprints into real and fake. Datasets from LivDet-2013, LivDet-2015 and LivDet-2017 were used to check the reliability of their proposed system. They managed to get the highest classification accuracy equal to 95.25% for LivDet-2017 dataset [44].
It can be seen that classifying fingerprints has almost been implemented by applying combinations of multiple processes, which may affect on their performances through spending long processing time. It can also be investigated from the literature that the accuracies of of detecting spoof fingerprints are not sufficient. We are aiming to contribute to this important area by presenting one intelligent model called the DFCN, which has the ability to detect the real and fake fingerprints with high performance and accuracy.

METHODOLOGY
This paper approahced an efficient deep learning model named the DFCN. This network model is based on the convolutional neural network (CNN). It aims to classify fingerprints into real (original) and fake fingerprints. Figure 2 shows the DFCN architecture. The proposed model consists of multiple layers. The following points summarize the layers of the DFCN model starting from a fingerprint input to the last classification layer: − Reading fingerprint image. The DFCN is adapted to accept an image of real or fake fingerprint in its input.
Any grayscale fingerprint image of size 300×300 pixels can be used as an input to our DFCN model. Each pixel value in a fingerprint image could be within the range [0-255]. − The convolutional process is the first layer in the DFCN. Two-Dimensional (2D) convolutions are here implemented for each input fingerprint image. We used 8 filters in this layer with the size 5×5 pixels. Each of the eight filters was specified to extract useful features from a fingerprint input. Let l is the current hidden layer, l-1 is the previous hidden layer and n is the number of nodes in the current layer. The following mathematical expression shows the general formula of the convolutional layer: where: is the n-th output of l layer, is the m-th input to l layer, , is the convolution kernel between the input and output (m-th and n-th, respectively), (*) is the operation of the convolution, −1 is the number of input channels, and represents the bias of the n-th output [45]. − Activation function is used after the convolution layer. The Rectified Linear Unite (ReLU) is employed as the activation function. The following formula represents the ReLU operation: where: is an output of the ReLU activation function [46]. − Max-pooling layer is applied to shrink the size of resulted data from the previous layer. The size of used filter in this layer was 5×5 pixels with the stride of 5 pixels. The max-pooling layer took the highest value of each 5×5 block of pixels. This value was considered as the most activated pixel on behalf of other pixels. This layer lead to decrease the computational load and the required processing time. It also plays a role in decreasing the overfitting process. The following formula represents the function of this layer: where: is an output of the max-pooling layer and Z is a block of 5×5 pixels. − Fully connected layer is provided to perform matching between the previous layer (the max-pooling layer in our case) and the next layer. Based on the function of this layer, the number of output classes can be specified here. In this study, 2 outputs are required for the real and fake classes. Then as mentioned, the fully connected layer will match between the number of required classes and the nodes of the max-pooling layer. − Softmax layer is subsequently given to prepare the data for the classification purposes. The output of this layer is represented by values in the range of 0 to 1. Each value gives the probability of the relationship between the current fingerprint input and an output class. The mathematical formula of the softmax function is represented as follows: where: is an output of the softmax layer and −1 is an input to this layer and C is the number of classes. The denominator is the summation of the exponentials of all the values in the vector [45]. − The final layer in the proposed DFCN model is the classification layer. It takes its inputs from the softmax layer as numbers between 0 and 1, two values for each input fingerprint image. The function of this classification layer is to decide whether the fingerprint is real or fake depending on the provided values from the softmax layer. The closer the input to 1 lead to select its corresponding output class and vice versa.

EMPLOYED DATASET
To assess our model, we have used two groups of real and fake fingerprint images from the ATVS-FakeFingerprint DataBase (ATVS-FFp DB), specifically, the without cooperation dataset. The middle and index fingers of both hands were used to capture their fingerprints by a capacitive sensor. Four fingerprint images for each finger were acquired [1,47]. The total numbers of fingerprints that are utilized in this study is 512 images. They are equally divided into two groups of real and fake fingerprints. We have exploited half of the dataset (50% of real fingerprints plus 50% of fake fingerprints) for the learning purposes. Whilst, the second half has been exploited for the DFCN efficiency testing.

RESULTS AND DISCUSSION
The proposed approach has been thoroughly evaluated by changing different deep learning parameters. Many experiments were implemented to examine the accuracies of tuning various deep learning parameters. It is worth mentioning that the starting parameters were initialized according to the novel approach in [48]. Table 1 shows the 8 cases that are considered in this paper, where each case represents tuning values for a certain parameter.
To reach the most efficient performance of the suggested DFCN model, the tuning processes of the deep learning parameters are divided into eight cases. Each of these eight cases includes adjusting one parameter while leaving the others without alteration. The number of epochs was assigned to 50 for all experiments as this would provide fair comparisons between all experiments and address the problem of implementing variations.
Starting with the initialization case, the proposed parameters of the novel approach in [48] were used without adjustment. These variables yielded the first accuracy of 87.89% in Table 1. The filter size in this case was modified in descending steps from 1515 pixels to 33 pixels while the other parameters were left unchanged. The highest accuracy in this case was 96.88% when using the filter size of 55 pixels. The filter size (55 pixels) that gave the highest accuracy in the first case was fixed in the second case whereas the number of filters was considered as the changeable variable from 16 to 4 filters. The best number of filters was 8 filters, where it achieved the highest accuracy of 99.22% and this is the rationale behind using this number of filters. Increasing or decreasing this number leads to reduce the performance of the applied application as it can clearly be seen in the same table. The rest cases of 3, 4, 5, 6, 7 and 8 in the same table included changing only one parameter in each case. However, the experiments showed that the alteration of the other parameters in these steps did not enhance the accuracy of the proposed model. process from the 0 padding, which gave the best results, to the 1, 2, 3 and (same) padding values. Case 5 examined the changing of pooling filter size, followed by case 6 which comprised of changing the pooling stride. The final two cases 7 and 8 included the adjusting of padding value and type of pooling, respectively. At the end of the tuning experiments, the parameters that have resulted with the highest accuracy are benchmarked for the DFCN model.  Figure 3 shows the accuracy of each case with respect to its changeable parameter. Additional executed experiments for assessing different deep learning optimizers. The three training optimizers: stochastic gradient descent with momentum (SGDM), adaptive moment estimation (Adam) and root mean square propagation (RMSProp) were evaluated according to the classification accuracy. Table 2 summarizes the effects of examining these optimizers on the DFCN accuracy.  Figure 3. The accuracies of each case with respect to its changeable parameter (the number of epochs was assigned to 50) Figure 4 shows the training curves of our suggested approach. The parameters that are utilized for training stage were set up as follows: SGDM optimizer, momentum=0.9, weight decay=0.0001, learning rate=0.03, minimum-batch size=128 and maximum epoch=50. The curves of the accuracy and the loss show that the suggested model started with high loss and low accuracy during the first iterations. However, the loss was then dramatically decreased and the accuracy was significantly increased at the same time. The curves show that the loss and accuracy reached their best performances and continued until the end of training.
For establishing fair comparisons with state-of-the-art networks, previous proposed models have been simulated and evaluated for our employed dataset and outputs. Three deep learning networks were used in these comparisons. These are the CNN [48], DFTL [49] and DFL [50]. Table 3 Table 3. Comparisons with state-of-the-art approaches Approach Accuracy (%) CNN [48] 90.23 DFTL [49] 92.19 DFL [50] 96.48 Our approach 99.22

CONCLUSION
Due to the importance of fingerprint as a physiological trait for individual recognition and due to the reality of its vulnerability to be breached by fake samples, we proposed an efficient deep learning model named the DFCN to classify fingerprints into two categories: real and fake. The suggested model was approached after applying many experiments that extensively evaluate various deep learning parameters. We used ATVS-FFp DB fingerprint images for our proposed model. The DFCN achieved great success with a classification accuracy reached to 99.22%. The experiments also showed that the DFCN outperformed state-of-art approaches when utilizing the employed dataset. In future work, different types of fingerprint fake samples can be considered. It appears that it is important to establish additional classification approach which can recognize the type of spoofing.