PHÁT HIỆN GIẢ MẠO KHUÔN MẶT SỬ DỤNG CÔNG NGHỆ TRÍ TUỆ NHÂN TẠO

Abstract

Fraud face detection is a crucial procedure for many face recognition systems.
In recent years, state-of-the-art approaches based on convolution neural networks (CNNs) show impressive results compared to traditional methods using hand-crafted features. In addition, the increasing trend o f embedding the computer vision systems on mobile devices requires that the designed algorithms are capable o f dealing with the time-critical constraint. In this paper, we first propose a CNN model, namely hduNet, developed from Google’s MobilenetV2 that provides a flexible trade-off between latency and accuracy, to detect different face spoofing attacks. We then provide an addition dataset o f roughly 5000 images capturing the characteristics o f Vietnamse people. Combining with LCC_FASD [1] dataset (which is only 1942 real face images, while having 16855 fake face images), the proposed model is carefully fine-tuned to optimize the computational cost as well as the classification accuracy.
To validate the model, different experiments have been conducted, demonstrating interesting performance in comparison with other methods.
Keywords: Face anti-spoofing, transfer learning, fine-tunning, convolution neural network.