A Cancellable Face Template Scheme Based on Nonlinear Multi-Dimension Spectral Hashing
Xingbo Dong, KokSheik Wong, Zhe Jin, Jean‐Luc Dugelay · 2019
The exposure of face templates potentially leads to severe security and privacy risks. For example, the attacker can utilize the compromised face template to masquerade the template owner. In addition, these concerns are aggravated since face is irreplaceable and irrevocable. In this paper, we propose a cancelable transform, namely nonlinear multi-dimension spectral hashing (NMDSH) to protect face template. Essentially, NMDSH utilizes a many-to-one function to transform real-valued deep face feature vector into binary code. The transformed template thus possesses strong non-invertible property. Next, a highly nonlinear softmod function is further adapted into the scheme to provide an additional layer of protection against similarity-based attack. The accuracy performance of NMDSH is evaluated. Experiment results suggest that NMDSH can preserve the accuracy performance largely. Properties including non-invertibility, revocability and resistance to similarity-based attack are also evaluated.