三维深度点云监督和置信度修正的人脸欺诈检测算法OACSCDCSTPCD

Due to feasibility and friendly user interaction,deep learning-based face recognition and identity authenti-cation becomes one of the most popular artificial intelligence technologies in China.The face recognition and identity au-thentication system should secure that the captured face for verification is a living face rather than a fake face or called spoofing face.Otherwise,the output of the system is useless for business.Face spoofing detection or called living face de-tection mechanism is set in the front part of the system,and plays a key role in distinguishing a fake face from the input fac-es.Most current face anti-spoofing algorithms perform well in intra-dataset.However,the model training in lab is unable to simulate all aspects in the real-world application scenarios.As a result,the data distribution in source domain is not always similar to the data distribution in target domain,which causes the lab-trained algorithms barely perform as well as in lab.Al-though we can mitigate the performance degradation with the increase of detection feature types and dimensions,it tends to make the detection network very complex in structure and large in model size.In order to improve the generalization ability of model without resorting to large model,we design a face spoofing detection network using the 3D depth point cloud su-pervision and confidence correction scheme.The proposed approach consists of three major contributions.First,we design a shallow convolutional neural network called DenseBlockNet.It can well extract distinctive depth features between real faces and spoofing ones and has a small model size.Second,we establish the relationship between the 2D depth map pro-duced by DenseBlockNet and the coordinates of sampling points,and thus create a 3D depth point cloud.We adopt the Chamfer loss to minimize the distance between the learned 3D depth point cloud and the ground truth 3D depth point cloud label,and use the binary cross entropy loss to supervise the difference between the learned 2D depth map and the ground truth 2D depth map label.Third,we introduce a prediction confidence map to correct the error of the learned 3D depth point cloud,so that it can avoid overfitting in intra-datasets and obtain good generalization ability in inter-datasets.Extensive ex-periments are conducted on 5 popular presentation attack databases,namely Reply-attack,CASIA-FASD,MSU-MFSD,Rose-Youtu,and OULU-NPU.Compared with 8 representative methods including 2 SOTA methods,the proposed method can achieve the least or second least half-total-error-rates in either intra-dataset or inter-dataset tests.Besides,it has the smallest model,the least amount of model parameters and the lowest computational complexity.