Face presentation attack identification optimization with adjusting convolution blocks in VGG networks

• DOI: 10.1016/j.iswa.2022.200107
• Publication Date: 2022-08-07T22:31:59Z
• Authors: Sudeep D. Thepade; Mayuresh Dindorkar; Piyush Chaudhari; Shalakha Bang
• Keywords: Liveness; Identification; Hyperparameter

Advancement in deep learning is mapping to every field of life and applying it almost all research problems. Numerous Deep Convolutional Neural Network (DCNN) architectures are being proposed, giving different results based on the depth value hyperparameters. The entire development such DCNN from scratch needs a lot effort, may not be used for other applications than one they structured for. Transfer way modify these pre-trained networks make them suitable newer diverse applications. This paper attempts empirically assess performance suitability existing human face liveness detection. Due advent ambient computing contactless identification humans using their biometric traits, detection proves an important area. Six models, alias VGG16, VGG19, DensNet121, Xception, MobileNet, InceptionV3, considered empirical assessment method explored two datasets - NUAA Replay-Attack. Face Liveness Accuracy, Half Total Error Rate (HTER) prime evaluation metrics. At rate 10−4, VGG19 network with scenario "Original VGG" gives highest accuracy, which outcome current research.