An enhanced light weight face liveness detection method using deep convolutional neural network

Abstract

Authentication plays a pivotal role in contemporary security frameworks, with various methods utilized including passwords, hardware tokens, and biometrics. Biometric authentication and face recognition hold significant application potential, albeit susceptible to forgery, termed as face spoofing attacks. These attacks, encompassing 2D and 3D modalities, pose challenges through fake photos, warped images, video displays, and 3D masks. The existing counter measures are attack specific and use complex architecture adding to the computational cost. The deep transfer learning models such as AlexNet, ResNet, VGG, and Inception V3 can be used, but they are computationally expensive. This article proposes LwFLNeT, a lightweight deep CNN method that leverages parallel dropout layers to prevent over fitting and achieves excellent performance on 2D and 3D face spoofing datasets. The proposed methods is validated through the Cross-dataset train test evaluation. The methodology proposed in the article has the following key contributions:

• •
  Design of Light Weight Dual Stream CNN architecture with a parallel dropout layer to minimize over fitting issue.
• •
  Design of Generalized and Robust deep CNN architecture that detects both 2D and 3D attacks with higher efficiency compared to existing methodology.
• •
  Method validation done with State-of-the-Art methods using the standard performance metrics for face spoofing attack detection.