id-Privacy in large scale biometric systems

Balancing privacy and security concerns in biometric systems is an area of growing importance. While important work has gone on in template protection and revocable biometric tokens, these avenues of research address only one aspect of the problem. Such research does not address a critical issue: balancing the need government and anti-fraud programs to do deduplication (ensure one identity per person) against the potential for abuse using that data. Any existing system capable of deduplication, even if using a template protection scheme, would allow function creep or abuse by searching with latent prints. This paper introduces the concept of id-privacy, requiring at least i items (e.g. fingers) to be provided to resolve identity to better than d above random chance. We show how using cross-finger representation on unsegmented fingerprint slap data, we can address what may be the single most important “privacy” issue in biometrics, privacy enhanced deduplication. We prove we can achieve (2,0)-id-privacy for fingerprint-based deduplication while preventing searching with a latent print. We introduce the Forest Finger algorithm - an approach for matching unsegmented slaps and cross-finger representations. Our results on the largest public slap database shows superior accuracy when compared with existing NIST Bozorth matcher when tested on unsegmented slaps, segmented prints or fused rolled prints.