Face Presentation Attack Detection Under Out-of-distribution Scenes

The task of face presentation attack detection,as a prerequisite for protecting face recognition systems,has become a widely studied topic by scholars both at home and abroad.Deep learning-based face presentation attack detection methods provide accurate and stable detection results when the test data and training data obey the same distribution.However,when there is a distribution shift between the training data and test data,the accuracy of the model decreases significantly,which poses great risks to face recognition systems.This thesis studies face presentation attack detection task for out-of-distribution scenarios,particularly addressing the problems of domain shift in test data and the inability of training data to contain attacks.The main research work is as follows:(1)For the task of domain adaptation,this thesis proposes a face presentation attack detection method based on domain invariant live features dual decomposing and progressive adversarial alignment.Firstly,this thesis heuristically decomposes the source domain features into domain specific features and domain invariant features.Then,the gradient of classifier is used to perform a second decompose of the live-related and live-unrelated features in the domain invariant features.This thesis adopts a curriculum learning approach to progressively align target domain features and the combination of live-related and live-unrelated features,gradually increasing the weight of live-related features and enhancing the correlation between the target domain features and face presentation attack detection task.From a causal perspective,this thesis provides an explanation for the domain-adaptive live alignment.The experimental results on four publicly available datasets show that the method proposed in this thesis compared with the existing seven methods achieves the current state-of-the-art performance on five testing protocols,and obtains the best HTER value,which reaches the current advanced level.(2)For the task of domain generalization,this thesis proposes a face presentation attack detection method that combines counterfactual intervention techniques in causal inference in order to strengthen domain related features.The hierarchical features are extracted by the backbone network to decouple shallow-level style information and deep-level content information.Specifically,this thesis uses a domain adversarial way to extract generalized domain-unrelated discriminative features for the content information while extracting domain-related features for the style information.This thesis introduces the causal effects before and after intervention into the loss function using counterfactual interventions to enhance the auxiliary role of domain-related information for the face presentation attack detection task.The experimental results on four publicly available datasets show that the method proposed in this thesis compared with the existing six methods obtains the best HTER and AUC value on CIM-O,which shows competitive generalization performance.(3)For the one-class face presentation attack detection task with only real face training data,this thesis proposes an end-to-end anomaly detection method based on a novel pseudo negative sample synthesis strategy.In order to improve the performance of the model in the face of unknown attacks,pseudo negative samples are synthesized in the low likelihood region of bonafide face feature space by learning the compact decision boundary between bonafide faces and unknown attacks,which can represent various unknown attacks in the training stage.To alleviate the positive and negative sample imbalance in the iteration process,this thesis uses focal loss as classification loss and pairwise confusion loss as regularization to train the final classifier.The intra-database testing results on the Idiap Replay-Attack and MSU-MFSD datasets demonstrate that the proposed method significantly reduces the ACER value compared to existing methods.