# US 2005/0238207 A1 · Biometric verification system and method utilizing a data classifier and fusion model

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2005/0238207 A1
- Document ID: US20050238207A1
- Status: Published application
- Kind: A1
- Application number: 10/831,374
- Filed: 2004-04-23
- Published: 2005-10-27
- Earliest claimed: 2004-04-23
- Assignees: Hitachi, Ltd.
- Inventors: Clifford Tavares
- Web page: https://patentgrep.com/patent/US20050238207A1

## CPC

- G10L 17/10: Multimodal systems, i.e. based on the integration of multiple recognition engines or fusion of expert systems
- G06F 18/256: {of results relating to different input data, e.g. multimodal recognition}
- G06V 40/10: Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G10L: SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06V: IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING

## Abstract

A biometric verification system and methods for using thereof are disclosed. The biometric verification system may comprise a verification engine for verifying an identity of a subject and a training engine. The verification engine may comprise a plurality of discriminators, a discriminator fusion component, and a subset fusion component. The discriminators analyze biometric inputs to generate a plurality of distortion scores. The discriminator fusion component analyzes the distortion scores utilizing a plurality of reference user models and a discriminator reliability ranking to obtain a combined confidence score for biometric sets of the subject in question. The subset fusion component analyzes the distortion scores for a plurality of biometric subsets utilizing subset selection-criteria to generate a combined confidence score for the biometric subsets of a subject in question. The verification engine may also include logic capable of determining whether to accept the subject in question as a genuine subject based on the combined confidence scores.

## Claims

### Claim 1

A biometric verification system, comprising:
a) a verification engine for verifying an identity of a subject, comprising:
i) a plurality of discriminators for analyzing biometric inputs of a subject in question to generate a plurality of distortion scores, the biometric inputs comprising biometric inputs for a plurality of biometric sets, each discriminator analyzing the biometric inputs of an associated biometric set, each biometric set being associated with a corresponding portion of the distortion scores;
ii) a discriminator fusion component for analyzing the distortion scores of the discriminators utilizing a plurality of reference user models and a reliability ranking of the discriminators to obtain a combined confidence score for the biometric sets of the subject in question;
iii) a subset fusion component for analyzing the distortion scores of at least one discriminator for a plurality of biometric subsets of the biometric set associated with the at least one discriminator utilizing subset selection criteria to generate a combined confidence score for biometric subsets of the subject in question; and
iv), logic for determining whether to accept the subject in question as a genuine subject based on the combined confidence scores; and
b) a training engine for training at least one discriminator, comprising:
i) training data for the biometric set associated with the at least one discriminator, the at least one discriminator analyzing the training data to generate a plurality of training distortion scores, the biometric set associated with the at least one discriminator comprising a plurality of biometric subsets, each of the biometric subsets having a portion of the training data associated therewith so that each biometric subset is associated with a corresponding portion of the training distortion scores;
ii) a score analyzer for analyzing the training distortion scores to generate reference user models for each of the biometric subsets and the biometric set of the at least one discriminator;
iii) a subset priority analyzer for analyzing the training distortion scores and reference user models generated by the score analyzer to develop subset selection criteria for the biometric subsets associated with the at least one discriminator; and
iv) a model analyzer for analyzing the reference user models generated by the score analyzer and the subset selection criteria generated by the subset priority analyzer to characterize the reliability of the at least one discriminator.

### Claim 2 (depends on 1)

The biometric verification system of claim 1, wherein each reference user model has a probability density function for genuine subjects and a probability density function for imposters that define a cumulative probability density function.

### Claim 3 (depends on 2)

The biometric verification system of claim 2, wherein the probability density functions for genuine subjects and imposters intersect at a point of indecision and have overlapping portions adjacent the point of indecision which define a region of indecision.

### Claim 4

A method for training a biometric verification system, comprising:
analyzing training data for a biometric set utilizing a discriminator to generate a plurality of distortion scores, the biometric set comprising a plurality of biometric subsets, each of the biometric subsets having a portion of the training data associated therewith so that each biometric subset is associated with a corresponding portion of the distortion scores;
analyzing the distortion scores to generate reference user models for each of the biometric subsets and the biometric set;
analyzing the distortion scores and reference user models to develop subset selection criteria for the biometric subsets; and
analyzing the reference user models and the subset selection criteria to characterize the reliability of the discriminator.

### Claim 5 (depends on 4)

The method of claim 4, wherein the reference user module for a biometric set is a cumulation of the reference user models of all of the biometric subsets associated with the biometric set.

### Claim 6 (depends on 4)

The method of claim 4, wherein each reference user model has a probability density function for genuine subjects and a probability density function for imposters that define a cumulative probability density function.

### Claim 7 (depends on 6)

The method of claim 6, wherein the probability density functions for genuine subjects and imposters intersect at a point of indecision and have overlapping portions adjacent the point of indecision which define a region of indecision.

### Claim 8 (depends on 7)

The method of claim 7, wherein the reliability of the discriminator is characterized as a discriminator reliability score, the reliability score being inversely proportional to the region of indecision.

### Claim 9 (depends on 4)

The method of claim 4, wherein the reliability of the discriminator is compared to the reliability of one or more additional discriminators to rank the reliabilities of the discriminators.

### Claim 10 (depends on 4)

The method of claim 4, wherein the analysis of the distortion scores and reference user models utilizes circumstantial data, demographic data, and statistical data to develop subset selection criteria for the biometric subsets.

### Claim 11

A method for verifying an identity of a subject, comprising:
analyzing biometric inputs of a subject in question utilizing a plurality of discriminators to generate a plurality of distortion scores, the biometric inputs comprising biometric inputs for a plurality of biometric sets, each discriminator analyzing the biometric inputs of an associated biometric set, each biometric set being associated with a corresponding portion of the distortion scores;
analyzing the distortion scores of the discriminators utilizing a plurality of reference user models and a reliability ranking of the discriminators to obtain a combined confidence score for the biometric sets of the subject in question; and
determining whether to accept the subject in question as a genuine subject based on the combined confidence score.

### Claim 12 (depends on 11)

The method of claim 11, further comprising analyzing the distortion scores of at least one discriminator for a plurality of biometric subsets of the biometric set associated with the at least one discriminator utilizing subset selection criteria to combine confidence scores of biometric subsets of the subject in question.

### Claim 13 (depends on 12)

The method of claim 12, wherein the subset selection criteria are derived from an analysis of reference user models in view of at least one of: circumstantial data, demographic data, and statistical data.

### Claim 14 (depends on 11)

The method of claim 11, wherein the discriminators are utilized in an order based on the reliability ranking.

### Claim 15 (depends on 11)

The method of claim 11, wherein the reference user model comprises a cumulative probability density functions for genuine subjects and for imposters that defines a point of indecision and a region of indecision adjacent the point of indecision.

### Claim 16 (depends on 11)

The method of claim 11, wherein the analysis of the distortion scores includes:
calculating an accept score and a reject score for the distortion score of each discriminator using the corresponding reference user model; and
determining whether the accept and reject scores for any of the discriminators lie in a region of indecision of the corresponding reference user model.

### Claim 17 (depends on 16)

The method of claim 16, wherein if the accept and reject scores of the discriminator having the highest reliability ranking lies outside the region of indecision of the associated reference user model, calculating the combined confidence score using the difference between the accept and reject scores of the highest reliability ranking discriminator.

### Claim 18 (depends on 16)

The method of claim 16, wherein if none of the accept and reject scores lie outside the regions of indecision of the corresponding reference user models, calculating the combined confidence score using the difference between the accept and reject scores of the highest reliability ranking discriminator and determining whether the combined confidence score is greater than zero.

### Claim 19 (depends on 16)

The method of claim 16, wherein if a portion of accept and reject scores lie outside the regions of indecision of the corresponding reference user models, calculating the combined confidence score using the difference between the accept and reject scores of the discriminator having the highest reliability ranking that lies outside the region of indecision of the associated reference user model.

### Claim 20 (depends on 11)

The method of claim 11, wherein the plurality of discriminators comprises a predetermined number of discriminators associated with a selected security level selected from a set of security levels.
