Research On Robust Face Spoofing Attack Detection

The increasing popularity and easy accessibility of face modalities have made face recognition systems(FRS)a major target of attack.Face spoofing attack(also known as presentation attack),as one of the most common and easy-to-implement attacks,can fool the face recognition system simply by presenting a face artifact(such as photo,video,or mask)of a legitimate user to gain unauthorized access to the biometric system.This will not only pose a serious threat to the security of face recognition systems,but also seriously harm the privacy and interests of legal users.Developing face anti-spoofing or presentation attack detection(PAD)methods(also known as liveness detection)to determine whether the face is real or fake is the efficient countermeasure.As the face spoofing uses non-living materials to make the attack,it will result in differences from the real faces in physiological features and vital signs.Therefore,face anti-spoofing techniques aim at extracting these differences to detect face spoofing attacks in FAS so that they can efficiently resist the increasing identity cheats.The research on face spoofing detection has been developing rapidly,but existing studies are faced with the following challenges: 1)there is no systematic evaluation or benchmark of the state-of-the-art face anti-spoofing methods on a common ground.Therefore,it is difficult to tell which methods perform better,especially in practical mobile authentication scenarios;2)the spoofing detection performance is generally affected a lot by the attack data,while most face spoofing attacks in existing datasets are of low quality and poor authenticity,and hard to realize the high data diversity in practical applications;3)existing methods may suffer from performance degradation on more diverse and realistic face spoofing attacks;4)most existing detection methods take antispoofing as a binary or one-class classification task,which not only ignore the differences among different attacks,but also have poor detection robustness on diverse or unknown spoofing attacks.This thesis aims at addressing the above problems in face anti-spoofing,and explores how to enhance the detection robustness,practicability,and generalizability of face antispoofing from the perspectives of performance evaluation,spoofing attack data collection,and detection algorithm design.The main research contents and innovations are summarized as follows:(1)Comprehensive evaluation on face anti-spoofing methodsTotally 20 representative face PAD methods have been evaluated on a common ground(i.e.,using the same datasets,same data preprocessing operations,same metrics,same protocols,and same classifiers).They have been evaluated on three public mobile spoofing datasets to quantitatively compare the detection performance,including accuracy,and robustness to diverse or unknown attacks.Meaningful observations and insights have also been summarized,showing the significant influence of spoofing attack data and feature extraction on detection performance.(2)Realistic face spoofing attacks based on wax figure facesConsidering the great influence of spoofing attack data on anti-spoofing performance,two hyper-realistic and diverse face spoofing attack datasets have been built.Based on the lifelike properties of wax figures,which have been used for identity fraud in reality,this thesis is the first work to introduce wax figure faces as one new type of super-realistic face spoofing attack.A wax figure face database(WFFD)with paired wax figure faces and real faces in both images and videos has been first collected from online resources,containing 2300 matched faces from 745 subjects and 285 videos from 241 subjects.The vulnerability of three popular FRSs and several representative face antispoofing methods to this kind of new attack has been investigated.Then a large-scale single wax figure face dataset(SWFFD)has been proposed with 4000 wax faces from1475 subjects to provide rich and highly-diverse data for face anti-spoofing.(3)Face anti-spoofing with factorized bilinear codingTo address the problems that various face anti-spoofing methods may suffer from performance degradation on databases with more diverse and realistic face spoofing attacks,this thesis proposes to construct discriminative,robust,and complementary skininspired features for face anti-spoofing in a fine-grained manner.A novel method based on factorized bilinear coding for multicolor channels has been designed and evaluated on several face spoofing databases to show its effectiveness,robustness,and interpretability.(4)Auxiliary multi-class classification based face spoofing attack detectionConsidering that existing research on face anti-spoofing are mostly based on binary or one-class classification task,and tend to experience performance degradation for diverse or unknown attacks,an auxiliary multi-class classification based face spoofing attack detection method has been proposed.The differences among real faces and various types of face spoofing attacks(including printed photo attack,replay video attack,3D masks,and wax figure faces)have been used to design a novel strong-training(multiclass classification)and weak-testing(binary classification)detection strategy.By learning the differences between face pairs based on the pair-wise learning scheme-Siamese Network,and the robustness of face anti-spoofing to unknown attacks can be improved.To sum up,this thesis focuses on the issues that current face anti-spoofing techniques will suffer from performance degradation on different or unknown types of face spoofing attacks.It has first conducted comprehensive performance evaluation on several state-ofthe-art face anti-spoofing methods to quantitatively compare the detection accuracy and robustness,and explore the significant influence of attack data and feature extraction methods on detection performance.For more challenging hyper-realistic face spoofing,two large-scale and diverse 3D face spoofing datasets with wax figure faces have been introduced as a new type of super-realistic face spoofing attack.Then a novel detection method based on factorized bilinear coding for multicolor channels has been proposed to enhance the detection robustness to unknown attacks.Finally,for various types of face spoofing attacks,an auxiliary multi-class classification based face anti-spoofing method has been proposed.This work addresses the limitations in face anti-spoofing from the perspectives of anti-spoofing performance evaluation,spoofing attack data collection,and anti-spoofing method design.It not only enriches the face spoofing attacks,but also improves the robustness performance of face anti-spoofing,which will have great significance in applying and spreading face anti-spoofing techniques for face recognition systems in practice.