Biometric untargeted attacks: A case study on near-collisions

Abstract

Biometric recognition systems are now integral to many authentication and identification processes, prompting the need to understand their resilience under various attack scenarios. In this work, we analyze the security of such systems against untargeted attacks, where an adversary aims to impersonate any user without focusing on a specific target. Assuming a minimal leakage model—where only a binary acceptance or rejection is revealed—we derive upper and lower bounds on the attack complexity as functions of the template size, decision threshold, and database size. Our contributions apply to templates following a uniform distribution, such as randomized biometric templates or those derived from high-entropy secret sources. Many biometric template protection schemes, such as BioHashing or random projection-based transformations, combine biometric data with a high-entropy secret (e.g., a password or token). This combination is designed to produce pseudo-random outputs, making the uniform distribution a reasonable assumption for the transformed template space. As a result, our analysis covers two-factor authentication systems where biometrics are combined with a stored random secret or strong password. We use probabilistic modeling to assess the theoretical security limits of such systems. We investigate two practical attack scenarios: naive outsiders submitting random guesses, and multiple simultaneous attackers increasing the overall trial rate. We also introduce the notion of weak near-collisions to evaluate the risk of mutual impersonation due to close templates in the database. Our theoretical analysis is validated on real biometric datasets (LFW and FVC) using transformation schemes such as BioHashing. Finally, we provide practical recommendations for configuring system parameters to mitigate untargeted attacks and near-collision risks.