Face Anti-Spoofing Using Convolutional Neural Networks

基於卷積神經網絡的人臉活體檢測

Student thesis: Doctoral Thesis

View graph of relations

Related Research Unit(s)

Detail(s)

Awarding Institution
Supervisors/Advisors
Award date13 Aug 2019

Abstract

Face anti-spoofing techniques, also known as face liveness detection, are one of the essential components of modern face-recognition based users authentication and access control systems. Without face anti-spoofing support, modern face-biometric based access control systems are vulnerable to a diverse range of spoofing attacks, ranging from simple face photographs to realistic 3D face masks. In the recent decade, a variety of approaches have been proposed for face anti-spoofing using convolutional neural networks (CNN), owing to its capability to learn dynamic and generalizable feature representations. However, the latent capabilities of CNN for face anti-spoofing applications have not been fully explored. As a result, there are still open issues that need to be addressed to improve the performance of face anti-spoofing applications using CNN, particularly in the case of unseen face Presentation Attacks (PA). In this work, firstly, a data-randomization based approach, is proposed for training a CNN from scratch for face liveness detection. Secondly, the adaptive combination of convolutional features learned from real-world face images and deep network generated face images (DNG), in CNN is explored for face liveness detection. Thirdly, the idea of single-camera based face liveness detection is extended to stereo-camera based face liveness detection using CNN. Finally, a combination of hand-crafted features like Histogram of Oriented Gradients (HOG) features, and Local Binary Patterns (LBP) features are utilized with deep convolutional-features to learn perturbed feature maps for face liveness detection.

This work first proposed an efficient strategy for training deep CNN classifiers for face liveness detection task, which utilized continuous data-randomization in the form of small mini-batches during training deep CNN classifiers on small scale face anti-spoofing databases. Unlike conventional data-randomization techniques utilized for CNN training, the proposed data-randomization strategy is based on randomly picking small mini-batches of face images from the whole database during training a CNN architecture. Experimental results on CASIA and Replay Attack databases have demonstrated the promising performance of the proposed method compared to state-of-the-art approaches for face liveness detection in both intra-database and cross-database scenarios.

Conventionally, classifiers designed for face liveness detection are trained on real-world images, where real-face images and corresponding face PA are very much overlapped. However, a little research has been carried out in the utilization of the combination of real-world face images and face images generated by deep CNN for face liveness detection. Therefore, this work further evaluated the adaptive fusion of convolutional-features learned by convolutional layers from real-world face images and DNG face images for face liveness detection. Additionally, an adaptive convolutional-features fusion layer is proposed that adaptively balance the fusion of convolutional-features of real-world face images and face images generated by deep CNN during training. Extensive experiments on the state-of-the-art face anti-spoofing databases, i.e., CASIA, OULU and Replay-Attack face anti-spoofing databases with both intra-database and cross-database scenarios indicate the promising performance of the proposed method on face liveness detection compared to state-of-the-art methods.

Current state-of-the-art face liveness detection methods are mostly single camera-based. However, a little effort has been carried out for the development of low-cost stereo camera-based face liveness detection system. Therefore, this work further proposed a low cost and reliable face liveness detection system that utilizes the stereo-pair face images with CNN for face anti-spoofing. For this purpose, the disparity-maps from the first convolutional layers of a CNN are computed, with stereo-pair face images as an input. Subsequently, the rest of the CNN layers are supervised using the computed disparity-maps. Additionally, a novel video-based stereo face anti-spoofing database with various face PA and different image qualities has been proposed. Experiments on the proposed stereo-pair face anti-spoofing database are performed using various test case scenarios. The experimental results indicate that our proposed system obtained significant results and have excellent generalization ability.

A novel approach is proposed for face liveness detection by perturbing convolutional feature maps of candidate convolutional layer, in a CNN, with perturbative weights. The perturbative weights are adaptive convolutional-weights that are learned from the weighted combination of convolutional feature maps of candidate convolutional layer, and hand-crafted features such as HOG and LBP and their combination. For this purpose, a special perturbation layer is designed to learn the adaptive perturbative weights during CNN training. The proposed perturbation layer is fully-convolutional and adds very little overhead to the total trainable parameters count. Extensive experiments with state-of-the-art CASIA, Replay Attack, and OULU face anti-spoofing databases show promising performance of the proposed method both in intra-database and cross-database scenarios compared to state-of-the-art face anti-spoofing approaches. The experimental results also highlight the attention created by the proposed perturbation layer in convolutional feature maps and its effectiveness in general for face liveness detection.