基于解耦空间异常检测的人脸活体检测算法

A Face Anti-spoofing Algorithm Based on Anomaly Detection in Disentangling Space

现有基于异常检测的方法大多仅利用活体样本进行单类建模,这样的特征用于活体检测的泛化能力强但准确率不高。而且,活体人脸特征单类建模并没有考虑活体人脸样本的多样性。活体人脸样本的不同身份、环境、采集设备等因素都会导致活体人脸的特征表达不紧凑,这样使得假体样本特征容易混入其中。为了解决以上两个问题,本文提出了一种基于解耦空间异常检测的人脸活体检测算法。本文设计了单中心对比损失,使得活体人脸特征在不限制假体人脸特征分布的情况下表达地更加紧凑。本文还对活体人脸进行了特征解耦,将其特征分为两个子空间:活体检测特征空间、活体无关特征空间。活体检测特征空间不受其他无关因素的影响,结合单中心对比损失来提高模型的泛化能力。库内实验和跨库实验共在5个数据集上与最新的方法进行了比较,在OULU-NPU数据集中,协议1相比于性能第2的模型错误率下降超过一半,最具挑战的协议4取得了仅3.3%的错误率;在SiW数据集的三个协议中也取得更低的错误检测率;在跨库实验中本文算法也表现出不错的泛化能力,尤其是在从重放攻击和打印攻击跨到3D面具攻击的跨攻击类型的测试中相比于性能第2的模型错误率下降5.41%。本文提出的人脸活体检测算法在检测性能和泛化性能上均优于其他先进方法,算法应对未知数据和新的攻击类型的能力有所提高。

Most of the existing methods based on anomaly detection only used live samples for one-class modeling,and such features had strong generalization ability for face anti-spoofing but low accuracy.Moreover,the one-class modeling of live face features did not consider the diversity of live face samples.The different identities,environments,collection equipment and other factors of live face samples led to the incompact representation of live face features,which makes the features of spoof samples easily mixed into them.In order to solve the above two problems,we proposed a face antispoofing algorithm based on anomaly detection in disentangling space.In this paper,a single-center contrast loss is designed to make the representation of live face features more compact without restricting the distribution of spoof features.We also disentangled the features of the live face and divide its features into two subspaces:the anti-spoofing feature space and the liveness irrelevant feature space.The anti-spoofing feature space was not affected by other irrelevant factors,and the single-center contrast loss was combined to improve the generalization ability of the model.The proposed method was compared with state-of-the-art methods in intra-database experiments and cross-database experiments on a total of 5 datasets.In the OULU-NPU dataset,the error rate of protocol 1 dropped by more than half compared to the model with the second performance,and the error rate of the most challenging protocol 4 achieved 3.3%.It also achieved a lower error detection rate among the three protocols in the SiW dataset.In the cross-database experiments,the algorithm in this paper also showed good generalization ability.Especially in the cross-attack type test from replay attack and print attack to 3D mask attack,the error rate dropped by 5.41%compared to the second-performing model.The face anti-spoofing algorithm proposed in this paper is superior to other methods in detection performance and generalization performance,and the ability to deal with unknown data and new attack types has been improved.