Biometric Identification Bias Evaluation Method
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Biometric identification functions for unidentified latent prints in UL databases face issues with high false rejection and acceptance rates, leading to biased results and excessive human examiner workload due to over-learning and misrepresentation during CNN training, particularly in the context of latent print variability and background confusion.
Innovation Solution
A method to evaluate and minimize the bias of biometric identification functions by calculating a bias score based on false match pairs, which involves determining the number of false matches for candidate and reference prints and using these scores to optimize the identification function, thereby reducing the Doddington risk and improving reliability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If biometric identification functions are optimized using CNN networks with machine learning, then matching performance improves, but false acceptances and false rejections increase due to over-learning and misrepresentation in training databases
Solution Approach 1:
The patent applies preliminary action by evaluating bias before deploying the biometric identification function in production. The bias evaluation process uses training databases to calculate bias scores that predict potential false acceptances, allowing optimization of the identification function before it is used for actual identification tasks. This preliminary evaluation prevents the system from proceeding with biased models that would cause excessive false acceptances.
Solution Approach 2:
The patent implements feedback by using bias evaluation results to iteratively improve the biometric identification function. The bias score calculated from false match analysis provides feedback about the model's tendencies, which is then used to adjust training parameters, select better training databases, or modify the identification algorithm to reduce bias and false acceptances in subsequent iterations.
2Reliability
If the biometric identification function returns more potential candidates to human examiners, then false acceptances are reduced, but human examiner workload increases excessively
Solution Approach 1:
The patent applies partial action by having the biometric identification function return only a moderate number of potential candidates to human examiners, rather than all possible matches or none at all. The bias evaluation helps determine the optimal threshold for returning candidates, balancing the need to reduce false acceptances with the practical constraint of examiner workload and available time.
3Adaptability or versatility
If biometric identification functions are trained on latent print databases, then identification capability improves, but bias increases due to variability and background confusion in latent prints
Solution Approach 1:
The patent converts the harmful effect of latent print variability and background confusion into a benefit by using these challenging examples in bias evaluation. The bias evaluation process specifically tests the identification function against latent prints with various degradation patterns, turning the previously harmful variability into a useful diagnostic tool that reveals bias and guides optimization to improve robustness.
Data Source
AI summary
An evaluation method of a biometric identification function implemented from a candidate biometric data set and a reference biometric data set, the biometric identification function, applied to a candidate biometric data item and a reference biometric data item, returning a match, the method comprising steps of:obtaining pairs of false matches, each pair comprising a candidate biometric data item and a reference biometric data item not associated with a same known individual;determining, for each candidate biometric data item in a pair, a first number associated with said item and equal to the number of pairs of false matches obtained comprising said item, and/or a second number associated with said reference biometric data item and equal to the number of pairs of false matches obtained comprising said item;calculating a bias score based on a maximum of the first numbers and/or based on a maximum of the second numbers.