Robust Face Anti-spoofing Algorithm With Convolutional Prototype Learning

Face anti-spoofing technology is the last line of defense for face recognition technology.It has extremely important security implications.Relevant data in practical application scenarios is complex,which is caused by the endless emergence of attack methods,various shooting scenes,and obvious different attack devices.The above factors have brought great difficulties to the existing face anti-spoofing technologies.They cause exist technologies to face many problems in practical scenes.In order to solve the above problems,I propose a robust face anti-spoofing algorithm based on convolution prototype learning.Besides,I design a domain adaptation strategy in few shot setting.The strategy aims to solve the issues in practical business scenarios.Firstly,I use prototype learning technology to assign multiple prototypes to model the distribution of complex data.Then,the loss function is designed based on the particularity of the face anti-spoofing task to improve the effectiveness of supervision.Compared with the traditional binary supervision method,the algorithm can constrain the variance in terms of intra-class and inter-class,thereby effectively improving the robustness of the model in complex scenarios.Besides,I design an adaptive prototype selection strategy to solve the difficulty of setting the number of prototypes.Improper setting of the number of prototypes will seriously restrict the performance of the algorithm.However,manually selecting the appropriate number of prototypes is tedious and expensive.The selection strategy I proposed can solve this problem well,making the model more applicable in practical application scenarios.In addition,in order to solve the domain adaptation scenario under the condition of few samples,which is common in actual business scenarios,the strategy I proposed can surpass the commonly used methods of fine-tuning models in terms of accuracy and cost.Finally,I verify the effectiveness of the above functional modules by designing scientific experiments.Experiments show that the proposed algorithm achieves great performance on multiple datasets.