Hiding phase-quantized biometrics: a case of steganography for reduced-complexity correlation filter classifiers

This paper introduces an application of steganography for hiding cancelable biometric data based on quad-phase correlation filter classification. The proposed technique can perform two tasks: (1) embed an encrypted (cancelable) template for biometric recognition into a host image or (2) embed the biometric data required for remote (or later) classification, such as embedding a transformed face image into the host image, so that it can be transmitted for remote authentication or stored for later use. The novel approach is that we will encode quantized Fourier domain information of the template (or biometric) in the spatial representation of the host image. More importantly we show that we only need 2 bits per pixel in the frequency domain to represent the filter and biometric, making it compact and ideal for application of data hiding. To preserve the template (or biometric) from vulnerabilities to successful attacks, we encrypt the filter or biometric image by convolving it with a random kernel which essentially produces an image in the spatial domain which looks like white noise, so essentially both the frequency and spatial representations will have no visible exploitable structure. We also present results on reduced complexity correlation filter classification performance when using biometric images recovered from stego-images.