Client-specific anomaly detection for face presentation attack detection

• Published in: Pattern recognition Vol. 112; p. 107696
• Main Authors: Fatemifar, Soroush, Arashloo, Shervin Rahimzadeh, Awais, Muhammad, Kittler, Josef
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.04.2021
• Subjects: Anomaly detection; Biometrics; Client-specific information; Deep convolutional neural networks; Face spoofing detection; Biometrics; Deep convolutional neural networks; Client-specific information; Anomaly detection; Face spoofing detection
• ISSN: 0031-3203; 1873-5142
• DOI: 10.1016/j.patcog.2020.107696

•We propose an anomaly-based face spoofing detection solution using representations derived by different CNN architectures.•By training the anomaly detection systems on genuine access data only, we avoid overfitting to any specific face spoofing attack data, and achieve improved robustness to novel types of attacks.•We investigate the merits of exploiting client-specific information in both, building anomaly-based spoofing detectors, as well as setting client-specific thresholds.•By conducting experiments on three benchmarking anti-spoofing datasets, we demonstrate that the proposed client-specific anomaly detection solution delivers superior performance compared to the state-of-the-art approaches in unseen attack scenarios. One-class anomaly detection approaches are particularly appealing for use in face presentation attack detection (PAD), especially in an unseen attack scenario, where the system is exposed to novel types of attacks. This work builds upon an anomaly-based formulation of the problem and analyses the merits of deploying client-specific information for face spoofing detection. We propose training one-class client-specific classifiers (both generative and discriminative) using representations obtained from pre-trained deep Convolutional Neural Networks (CNN). In order to incorporate client-specific information, a distinct threshold is set for each client based on subject-specific score distributions, which is then used for decision making at the test time. Through extensive experiments using different one-class systems, it is shown that the use of client-specific information in a one-class anomaly detection formulation (both in model construction as well as decision boundary selection) improves the performance significantly. We also show that anomaly-based solutions have the capacity to perform as well or better than two-class approaches in the unseen attack scenarios. Moreover, it is shown that CNN features obtained from models trained for face recognition appear to discard discriminative traits for spoofing detection and are less capable for PAD compared to the CNNs trained for a generic object recognition task.