| Face verification is an identity authentication technology based on facial biometric characteristics,and it is widely deployed in many applications,such as train stations,banks,and online account login.Meanwhile,the security issues of face verification are also of concern,and its vulnerability to non-intrusive threats was demonstrated over recent years.Typical threats include presentation attacks(a facial artefact is presented to the capture subsystem)and morphing attacks(a morphed face image containing multiple subjects’ biometric characteristics is used in enrollment).To detect such attacks and to strengthen the security of face verification systems,we investigate the forensics and anti-forensics methods of face presentation attacks(PAs)and morphing attacks(MAs).The main contributions are as follows.1)A detection method based on color distortion and ensemble learning is presented for print and replay PAs.The existing work does not fully consider the correlation between different color channels as well as the optimization of classification for detecting PAs.To address these limitations,a color distortion-based detection model is first built,and then the chromatic discrepancies between bona fide faces and artefacts are extracted by cross-channel local binary patterns.Meanwhile,an ensemble learning-based classifier is put forward to reduce the effect of class imbalance and to improve the generalization ability.The experimental results and analysis show that it achieves good results in the cross-dataset test and can improve classification performance.2)A detection method using local face differential is proposed for adversarial facial accessory PAs.It extracts the local face differential features from a suspect face image and a reference face image,and then adaptively fuses the differential features of different local face regions for detection.Meanwhile,its principle is explained by theoretically investigating the local facial similarity.To evaluate the proposed method,a dataset is built,5 testing protocols and 21 different training/testing sets are provided.The experimental results indicate that the proposed method can effectively distinguish between adversarial facial accessory PAs and bona fide faces,and it has good generalization ability to unseen data.3)A detection method based on face embedding supervision is proposed for MAs.In the prior work,the classical score-level loss functions ignore the properties of different MAs,and they also cannot be directly applied to differential detection of MAs.To this end,fine-grained classification loss and differential loss are respectively devised based on the properties of different MAs and differential detection scenarios.The experimental results and analysis show that the former can further locate the morphed areas of a detected morphed image,and the latter can achieve improvement in the generalization to unseen MAs and the robustness to low-quality probe images.4)An empirical study on a novel and effective face impostor PA is made.In the proposed PA,a facial artefact is created by using the most vulnerable facial components,which are optimally selected based on the vulnerability analysis of different facial components to impostor PAs.An attacker can launch a face PA by presenting a facial artefact on his or her own real face.With a collected dataset containing various types of artefacts and PA instruments,the evaluation results show that the proposed PA poses a more serious threat to face verification and PA detection systems compared with the print,replay,and mask PAs.Moreover,the generalization ability of the proposed PA and the vulnerability analysis with regards to commercial systems are also investigated.5)A partial face manipulation-based MA is developed.It changes MA from a holistic face level to component level,and only the most effective facial components(eyes and nose)are used.Thus,a manipulated face is more similar to a bona fide one in terms of visual quality,texture,and noise characteristics.Moreover,a novel metric called actual mated morph presentation match rate is proposed to evaluate MA performance under real-world conditions.With a collected dataset containing different MA types,image qualities,and manipulation parameters,the results indicate that the proposed MA has better anti-detectability compared with the existing complete,splicing,and combined MAs.Moreover,it has low distortion and can reach a better tradeoff among facial biometrics verification,anti-detectability,and distortion.The proposed methods have the potential to be applied to face verification systems as countermeasures for PAs and MAs,and can improve the security of face verification to a certain extent. |
