Research On Face Liveness Detection Based On Deep Metric Learning

In recent years,with the rapid development of mobile internet,face recognition technology has been widely applied in various fields such as finance,security,and social networks.However,as the application scenarios expand,the systems face increasing challenges and issues.Traditional face recognition systems are vulnerable to attacks such as printed photos,face video playback,and 3D disguises,which pose security concerns.To enhance the security of face recognition systems,face liveness detection models have been introduced to determine whether the captured face belongs to a real living person or is a forged face attack.Face liveness detection is an evolving problem where attacks and defenses iteratively evolve,presenting constant challenges.Firstly,traditional deep feature-based binary classification methods tend to group different types of fraudulent samples into the same class,causing feature overlap and a decrease in accuracy.Secondly,as face liveness detection technology is applied in different domains,the collected datasets exhibit significant variations across different scenarios,posing challenges to the model’s generalization capability.To address these issues,this study proposes an improved face liveness detection approach that incorporates techniques such as metric learning and domain generalization to enhance the accuracy and generalization capability of the model.The specific contributions are as follows:(1)Introducing a metric learning approach based on cross-batch storage mechanism for face liveness detection.To tackle the problem of poor generalization in liveness detection,an anomaly detection method is employed to learn a compact representation space for liveness samples.Additionally,fine-grained fraudulent features are obtained through pixel-level auxiliary supervision.To address the issue of feature overlap between classes,metric learning is utilized to encourage intra-class compactness and inter-class separation.Sample mining strategies are improved by combining intra-batch and inter-batch sample mining,expanding the scope of sample mining to acquire more effective sample pairs.Furthermore,a multi-scale triplet loss is designed to optimize the model and achieve clearer classification boundaries.The proposed approach is validated through comparative experiments on publicly available datasets such as OULU,Replay Attack,and CASIA,demonstrating its robustness and generalization capability.(2)Proposing a cross-domain face liveness detection approach based on asymmetric focal triplet loss.Firstly,an improved Res Net18 network is designed to extract global features of samples,and domain adversarial training is employed to learn a shared feature space that aligns the feature distributions across source domains(multiple datasets).Secondly,the metric learning module incorporates the asymmetric focal triplet loss constraint,which maps the Euclidean distance to an exponential kernel and automatically increases the weight of challenging samples.This encourages compact intra-class distribution and dispersed inter-class distribution,forming better class boundaries.The classification module employs R-drop regularization technique to further improve model robustness and stability.The effectiveness of the proposed approach is verified through comparative experiments on publicly available datasets.