Research On Masquerade Attack On Biometric Hashing Via Generative Adversarial Network

With the widespread application of biometric recognition and the consequent security issues of biometric template,biometric hashing as an effective biometric template protection method has gained more and more attention in recent years.The hashing transformation of biometric hashing transforms a biometric image into a binary representation through a hashing mapping and the process is generally considered to be secure,and hence the security of biometric template is therefore guaranteed.The masquerade attack on biometric hashing as an attack which inverts the original biometric image from the given hashcode,has been given much attention recently.This attack is mainly used to validate the security of biometric recognition system or to enlarge the existing biometric database like face or iris.In the attack,some adversaries have tried to reconstruct original biometric images from transformed hashcode utilizing similarity preserving property on which its utility depends.However,an existing state of art method tends to ignore the perceptual similarity of synthesized biometric images in the attack and consequently the synthetic images can be easily differentiated from real images.To obtain the image which can simultaneously pass the validation of recognition system and maintain moderate perceptual similarity to real biometric image,we propose a novel masquerade attack network named Biohash GAN which is based on the generative adversarial network.The main contribution of our work are as follows:(1)We introduce a new target combining semantic invariability in hashing space and perceptual similarity in biometric space.For the semantic invariability,the hamming distance between the hashcode of original image and that of synthetic image is used to guarantee that the two images should be classified into the same category by the recognition system.And for the perceptual similarity,divergence between distributions of synthetic images and real images is used to guarantee that the synthetic image should look similar to the real image.(2)We propose a DNN-based network named hashnet to simulate the unknown and non-linear hashing mapping from biometric images to hashcodes.The hashnet is then used to tackle the derivative problem related to discrete hash mapping in hashing space.(3)We incorporate the hashnet into a generative adversarial network to constitute our reconstruction model named Biohash GAN.The generative adversarial network is to learn the mapping from image to hashcode so as to maintain the perceptual similarity and the hashnet is to regulate the generation process of generative network in the hashing space so as to maintain semantic invariability.Experiment result on dataset CASIA-Iris V4.0-Interval,CMU PIE and biometric hashing method Iterative quantization,Biohashing demonstrates that the synthetic images obtained from the proposed network Biohash GAN can pass the validation of recognition system and simultaneously maintain moderate perceptual similarity to real biometric images.