B0636
Title: Assessing the invertibility of deep biometric representation
Authors: Yiwei Wang - University of New South Wales (Australia) [presenting]
Clara Grazian - University of Sydney (Australia)
Abstract: Biometric recognition is increasingly applied in practice given the worries about security problems. Although various approaches have been used to protect databases of biometric systems, there still exists the possibility that the template of a person is stolen. Since biometric traits are physically connected to a person, we focus on the ability of networks to avoid the attach of using inverse imagining from the deep representation. We use convolutional neural networks to generate the deep representations and a GAN to reconstruct the image. Then we analyse the invertibility of the representation through random forest regression. We try to understand what features of the NN have the largest impact on the representation invertibility.
