Core Concepts
A secure and private biometric template matching system that uses multiple independent embeddings stored in separate vaults with chaff points to protect user identities and enable efficient authentication.
Abstract
The paper proposes a secure and private biometric template matching system that addresses the security and privacy concerns of traditional biometric systems. The key ideas are:
Divide the original biometric template into multiple sub-templates and store them in separate vaults, each with a large number of randomly generated "chaff" points to obfuscate the real template.
Use multiple independent deep learning models to generate distinct embeddings for each sub-template, making it computationally infeasible for an attacker to recover the original template.
During verification, the query is matched against the top vectors from each vault, and a successful match requires satisfying a threshold number of vaults. This improves the true positive rate while maintaining a low false positive rate.
The system leverages generative adversarial networks (GANs) to create realistic synthetic face images as chaff points, further enhancing privacy by avoiding the use of real user data.
Extensive experiments on the AT&T, Georgia Tech, and LFW face datasets demonstrate the effectiveness of the proposed approach, achieving high accuracy (AUC > 0.99) while providing strong security guarantees against brute-force attacks.
The system is computationally efficient, with the complete end-to-end process taking only around 1.5 seconds on average, making it practical for real-world deployment.
Stats
The paper reports the following key statistics:
The area under the curve (AUC) for the AT&T dataset is 0.9939.
The AUC for the Georgia Tech dataset is 0.9942.
The AUC for the LFW dataset is 0.9042.
The true negative rate (TNR) is 100% for both the AT&T and Georgia Tech datasets.
The true positive rate (TPR) ranges from 81.20% to 96.19% depending on the number of chaff points and the number of classifiers used.
Quotes
"Our approach leverages the power of generative models to create synthetic facial images, demonstrating their potential in building robust and secure end-to-end biometric systems."
"Our work offers distinct advantages as it eliminates the risk of exposing real individuals' identities during the system's training and operation."