Research On Face Anti-spoofing With Feature Fusion In Diverse Scenes

Face recognition has become a popular technology for various applications due to its accuracy and non-contact.However,this technology also brings some security concerns.Face images and videos are more readily accessible and less private compared to other biometric features like fingerprints and iris scans,making them vulnerable to unauthorized access.Malicious actors can acquire the facial information of others through a variety of methods,to bypass face recognition systems.This issue not only disrupts the proper functioning of the identification system but also jeopardizes personal privacy and interests.Therefore,verifying the authenticity and reliability of face information is crucial for the successful implementation of face recognition technology.Despite rapid advancements in face anti-spoofing technology,there are several issues that need to be addressed.For instance,the current face anti-spoofing models are susceptible to disturbance by complex and diverse backgrounds,making it challenging to detect spoofing face images effectively.Additionally,it is difficult for the model to obtain high-quality face features due to various factors that may influence the face images collected by different devices.Lastly,the challenges faced by the current technology are increasing seriously as the type of attack is diverse,but the available data to train the model is limited.To address these issues,this thesis proposes the following three methods to improve face anti-spoofing model in terms of robustness and generalization.(1)To prevent the model from being easily disturbed by complex and diverse backgrounds,a novel face anti-spoofing model based on facial texture information and fore-background difference analysis is proposed.The model consists of a Conv Ne Xt-based facial texture analysis module,which provides effective facial information for the detection model,and a forebackground difference analysis module,which makes the sample texture features richer.To ensure the high accuracy of the model in different environments and combine the advantages of the features from the above two modules,an attention feature fusion module is designed and improves the model’s performance more comprehensively.(2)To assist the model in obtaining high-quality face features,a face anti-spoofing method that integrates high-response topological graph structure features is proposed.This method constructs a topological graph structure feature extraction backbone network to assist the model in extracting long-distance dependent auxiliary information.A feature alignment strategy is designed to suppress redundant information in the graph structure features to improve the high responsiveness of the graph structure information with the original backbone features.In addition,to refine the extracted sample features,this method establishes a group perception field composite module for fusing dual-stream features to effectively improve the convergence speed and detection performance of the model.(3)To improve the ability of the face anti-spoofing model to process unknown scenarios and attack methods,this thesis designs a dual-stream face anti-spoofing model based on visual attention and domain feature fusion,applying domain generalization to face anti-spoofing tasks.A feature extraction module based on visual attention is proposed to enhance the model’s ability to capture content features.Additionally,adversarial training is used to prevent the model from focusing on distinguishing the domain of the sample,reducing the sensitivity of the model to domain features.Moreover,a novel contrast loss strategy is proposed to avoid the problem of training oscillation,further improving the model’s robustness.In summary,this thesis proposes three optimized methods to address the shortcomings of existing methods in terms of robustness and generalization.These optimized methods are conducted in three aspects: basic components of feature extraction,backbone network,and supervised learning methods.The proposed methods are evaluated via multiple performance metrics based on four datasets: CASIA-FASD,Replay-Attack,MSU-MFSD,and OULU-NPU.The experimental results demonstrate the significant advantages of the proposed methods over similar state-of-the-art methods,which have important implications for the practical application and promotion of face anti-spoofing methods.