ROLE OF DEEP LEARNING APPROACHES IN BIOMETRICS
Deep learning-based models have shown outrageous performance on a variety of application areas like computer vision (self-driving cars, robotics), natural language processing (automatic text generation, automatic machine translation), biomedical, forensics, and many more in recent years. Even the biometric recognition system does not remain untouched to deep learning. In recent years, deep learning-based models are extensively used in the biometric domain to improve the accuracy of different recognition systems based on various biometric traits like face, iris, fingerprint, palm- print, gait, and many more. There is a wide range of applicability of deep learning concepts in the biometric domain that ranges from segmentation, authentication to generation of artificial real looking biometric samples. One of the fundamental problems with handcrafted features in the biometric domain is that they are highly trait- specific (like LBP, Gabor-based features well suited for the face, log Gabor features for iris) and require parameter tuning according to the dataset in consideration. On the other hand, deep learning-based methods provide an end-to-end learning framework that automates the process of learning the best feature representation irrespective of biometric trait and thus more universal. It should be noted that the tremendous success of deep learning mainly leverages the availability of large datasets. Non-availability of labeled voluminous datasets in the biometric domain except for face is a fundamental problem. To circumvent this situation, the transfer learning paradigm can be used to handle labeled data constraint. Thus, in the literature [84,92], many deep learning- based models for biometric authentication are based on transfer learning paradigm.
2.3.1 Deep Learning in Face Recognition
As compared to other vision community datasets, there is large intra-class variability and high inter-class similarity in biometric datasets; thus, the applicability of deep learning models in the biometric domain is quite challenging. To handle this situation, many researchers have proposed domain-specific loss functions that deal quite well with intra-class dissimilarity and inter-class similarity. For example, in the year 2016, Liu et al. [55] proposed an alternative of cross-entropy loss by the name large margin Softmax loss, and claimed that this loss explicitly enhanced inter-class separability and at the same time encouraged intra-class compactness between learned features. They experimentally verified their claim on the Labeled Faces in the Wild (LFW) (face dataset) and achieved good results. In another notable work, Liu [54] proposed hypersphere embedding for face recognition by the name SphereFace. This hypersphere embedding is constructed by proposing a novel loss function (angular softmax) that helps convolutional neural network (CNN) in learning angular discriminative features. CosFace [98], UniformFace [22], ArcFace [18], and AdaptiveFace [53] are popular deep learning-based frameworks for face recognition. It should be noted that deep learning-based frameworks in the biometric domain are mostly employed in the face recognition domain due to the availability of large labeled face datasets; Ref. [99] gives a detailed overview of various deep learning- based face recognition techniques.
2.3.2 Deep Learning in Iris Recognition
Deep learning-based framework is mostly used in the iris domain for iris recognition and spoof detection. Traditional iris recognition mainly comprises three major steps:
- (i) iris segmentation, (ii) iris normalisation, and (iii) iris feature extraction and matching. It is worth mentioning that earlier works in iris recognition using deep learning- based techniques mainly focus on the iris feature extraction part only [9,27]. Like in the work carried out by Minaee et al. [57,58], they have extracted discriminative iris features from normalised iris images using the transfer learning paradigm on VggNet and ResNet models. Recently, it has been pointed out by Ahmad et al. [4] that due to the highly discriminative learning capability of deep networks, outrageous results on iris recognition can be achieved by directly feeding segmented iris images to deep networks without normalising them. They have experimentally validated their claim on several challenging iris datasets (ND-0405, UbirisV2, and IITD). Very recently, a unified framework (deep learning-based) that detects, segments, and recognises iris images simultaneously without any need for pre-processing was proposed by Zhao et al. [107]. Recently, GANs are also used for augmenting iris datasets. It has been pointed out by researchers [50] that traditional iris augmentation techniques result in generating highly correlated samples, and thus, they are not as robust as compared to the one generated through generative models. Through the progress made in iris recognition, one can see how biometric researchers are adapting deep learning frameworks for challenging biometric problems.
- 2.3.3 Deep Learning in Fingerprint Recognition
Deep learning-based approaches are widely used in fingerprint recognition. For example, FingerNet [91], a deep learning-based model for fingerprint minutiae detection, jointly performs tasks like feature extraction, segmentation, and orientation estimation; similarly in another work by Stajanovic et al. [86], CNNs are used for fingerprint ROI segmentation. There has also been some work for detecting fingerprint spoofs using deep learning-based techniques [45,65].
2.3.4 Deep Learning in Other Biometric Traits
Deep learning-based frameworks are not so popular for recognising other biometric traits apart from the face, iris, and fingerprints due to the non-availability of labeled training datasets as well as due to their nonpopularity in comparison to face, iris, and fingerprints. But still, biometric researchers have used deep learning frameworks by utilising the concept of few shot learning where deep models can be trained by using few training samples like the one done in palmprint recognition [81]. In this work, features extracted through CNNs are represented as nodes in graph neural networks. Apart from that, Siamese networks are also used for learning intra-class similarity and inter-class dissimilarity from small training datasets.
In addition, many works have been carried out so far [70,93] using deep learning- based techniques for mitigating presentation attacks in biometric traits, but preventing adversarial attacks and template attacks through deep learning techniques is still in their infancy state. We will discuss more on mitigating template attacks based on the deep learning-based architecture in the following sections.
