On the Practicality of Local Ranking-Based Cancelable Iris Recognition

Practical cancelable biometrics (CB) schemes should satisfy the requirements of revocability, non-invertibility, and non-linkability without deteriorating the matching accuracy of the underlying biometric recognition system. In order to bridge the gap between theory and practice, it is important to prove that new CB schemes can achieve a balance between the conflicting goals of security and matching accuracy. This paper investigates the security and accuracy trade-off of a recently proposed local ranking-based cancelable biometrics (LRCB) scheme for protecting iris-codes. First, the irreversibility of the LRCB is revisited and a new attack for reversing the LRCB transform is proposed. The proposed attack utilizes the distribution of order statistics for discrete random variables to reverse the protected rank values and obtain a close approximation of the original iris template. This ranking-inversion attack is then utilized to realize authentication, record multiplicity, and correlation attacks against the LRCB transform. Our theoretical analysis shows that the proposed reversibility attack can recover more than 95% of the original iris-code bits and the proposed correlation attack can correctly correlate two templates 100% of the time for the parameter setting that retain the recognition accuracy of the underlying iris recognition system. The validity of the proposed attacks is verified using the same iris dataset adopted by the authors of the LRCB scheme. Experimental results support our theoretical findings and demonstrate that the security properties of irreversibility and non-linkability can only be fulfilled at the expense of significant and impractical degradation of the matching accuracy.