Performance generalization in biometric authentication using joint user-specific and sample bootstraps

Area of Science:

• Computer Science
• Biometrics
• Machine Learning

Background:

• Detection error trade-off (DET) curves are standard for visualizing biometric authentication performance.
• Existing methods for evaluating DET curves are sensitive to database characteristics like sample selection, demographics, and user count.

Purpose of the Study:

• To develop a robust method for predicting biometric authentication performance (DET curves) that accounts for database variability.
• To enhance the reliability of DET curve analysis by incorporating sample selection, demographic composition, and user number.

Main Methods:

• A novel two-step bootstrap procedure was proposed, extending Bolle et al.'s technique.
• The method addresses variability from sample choice, demographic composition, and the number of users in a biometric database.
• Experiments were conducted on the NIST2005 and XM2VTS benchmark databases.

Main Results:

• Preliminary experiments show encouraging results, with an average of over 75 percent DET coverage when predicting unseen DET curves with eight times more users on NIST2005.
• The proposed bootstrap procedure effectively accounts for the specified sources of variability.
• Increased data availability leads to smaller and more informative confidence intervals for DET curves.

Conclusions:

• The proposed two-step bootstrap procedure offers a more reliable way to assess and predict biometric authentication performance.
• This method enhances the generalizability of DET curve analysis across different database compositions.
• The findings suggest improved confidence in performance predictions as more data becomes available.