On the Reversibility of Locality-Sensitive Hashing-based Biometric Template Protections

Seunghun Paik, Chanwoo Hwang, Sunpill Kim, Jae Hong Seo

Abstract: With the extensive deployment of biometric authentication systems, the need for biometric template protection (BTP) has been widely recognized. Designing a secure yet efficient BTP is still a long-lasting challenge, and locality-sensitive hashing (LSH) is a promising building block for designing BTPs. In this study, we propose a novel pre-image attack applicable to lots of existing LSH-based BTPs, showing that many of them are in fact reversible. Our attack leverages structural properties shared by several LSH-based BTPs. Through investigation, we formalize a certain class of LSHs vulnerable to our attack, called PMA-LSH, which contains several known LSH-based BTPs, even not-yet-cryptanalyzed ones. Furthermore, the recovered pre-image from our attack is much closer to the original template compared to previous attacks, facilitating recovery of the original biometrics via reconstruction attacks. With existing reconstruction methods, we successfully recovered biometrics from templates protected by several LSH-based BTPs. The recovered biometrics sufficiently resemble the original ones, so they can be further exploited to impersonate other recognition systems, including commercial APIs. To facilitate further study, our source code is publicly available at https://github.com/Cryptology-Algorithm-Lab/Analysis_LSH.

External IDs:doi:10.1109/tdsc.2025.3637307