Anomaly Detection in Face Anti-spoofing: Algorithms, Training Set Construction, and Bias Analysis

Abstract

Face recognition is a mature and trustworthy method for identifying individuals. Thanks to the availability of high-definition cameras and accompanying devices, this particular biometric recognition modality is widely regarded as the fastest and least obtrusive option. Despite advancements in face recognition systems, it has been discovered that successful spoofing attempts are still possible. Various anti-spoofing algorithms, also known in the literature as liveness detection tests and presentation attack detection algorithms, have been devised to counteract such attacks. The first contribution of this research is to demonstrate the effectiveness of certain simple and direct spoofing attacks. Our approach involves utilizing ResNet50, a highly reliable deep neural network, as a binary classification method. We assess its performance by subjecting it to adversarial attacks that involve manipulating the saturation component of imposter images. We have found that it is particularly vulnerable to spoofing attacks that employ processed imposter images. To the best of our knowledge, this study represents the pioneering exploration of adversarial attacks on deep neural networks within the realm of face anti-spoofing detection. In addition, we conducted an experiment that revealed the potential of the proposed adversarial attack to be converted into a direct presentation attack. In a second contribution, we propose an alternative approach incorporating in the- wild images and non-specialised databases into anomaly detection to improve the face anti-spoofing algorithm’s performance on unseen databases. We developed a method for detecting anomalies in face anti-spoofing by employing a convolutional autoencoder. We assessed its effectiveness using the NUAA database, which had not been previously utilized in the training. Our results indicated improved performance when incorporating in-the-wild face images and face data from nonspecialized databases into the training dataset. Transformers are emerging as the new gold standard in various computer vision applications and have already been used in face anti-spoofing, demonstrating competitive performance. In a third contribution, we propose a network with the ViT transformer and ResNet18 as the backbone for anomaly detection in face anti-spoofing with a decoder as the head. Then, we validate various anomaly detectors to compare the results with our proposed method. Also, using the ViT with MLP as a binary classifier baseline and compare it with our model. Our comprehensive testing and evaluation have demonstrated that this proposed approach competes admirably as a method for detecting anomalies in the domain of face anti-spoofing. Finally, there are only a few papers that specifically address the issue of racial bias in anti-spoofing. As a fourth contribution, we present a systematic study of race bias in face anti-spoofing with three key characteristics: the focus is on analysing potential bias in bona fide errors, where significant ethical and legal issues lie; analyses of various stages of the classification process, and treating the value of the threshold that determines the classifier’s operating point on the ROC curve as a user-defined variable. We do not assume it is fixed by the vendor of the biometric verification system through a black-box process. To the best of our knowledge, this is the first investigation into racial bias within the face anti-spoofing domain that employs anomaly detection techniques while also incorporating a non-specialized database for analysis. Our results show that racial bias in face anti-spoofing is influenced by factors beyond mean response values, such as different variances, bimodality, and outliers. Overall, this thesis contributes to the ongoing development of anti-spoofing techniques and investigates some important issues regarding the potential for bias in these systems.