Presentation attack on face biometrics

Abstract

There is a rising interest in reducing the cost and enhancing the precision of the design of deep network topologies. Traditional authentication systems which include pin/password based login is the most often used regular login authentication technique that you will use on a daily basis when using an online service. When utilizing the Password-Based Authentication approach, you must input your username or mobile number as a user id and a password. Only once both of these aspects have been confirmed is the individual approved. HydraNets are large networks that include unique components that are specialized to calculate features for visually comparable classes. However, in order to maintain efficiency, these networks dynamically pick just a limited number of components to assess for each individual input picture. This design is made feasible by a soft gating mechanism that supports component specialization during training and properly conducts component selection during inference. Specifically, this architecture encourages component specialization during training. On the NUAA, CASIA and OULU-NPU classification problems, the HydraNet technique is evaluated and analyzed here. When applied to the ResNet50 and DenseNet121 models at above mentioned datasets, the HydraNet framework has been shown to lower the cost of inference by 2-4x while maintaining the accuracy of the baseline architectures. Keywords: Hydranet, Resnet50 and DenseNet121