Described herein are a system and techniques for identifying and preventing certain fraud attacks that may be used to defeat facial recognition systems. In embodiments of the system described herein, biometric data may be segregated into regions, which are then processed separately and in parallel. Likeness scores are determined for each of the separate regions. By tracking individual region likeness scores used in access requests in accordance with embodiments of the disclosure, the system is able to identify potential fraud attacks that cannot be detected using conventional systems.