Granted Patent B1

US 12,315,294 · App. 17/725,978 · Granted May 27, 2025

Inventors: Norman Hoon Thian Poh (Atlanta, GA); Ramprakash Srinivasan Puri (Atlanta, GA); Luke Arpino (Atlanta, GA); Daryl Burns (Atlanta, GA)

Assignee: T Stamp Inc.

G06V40/168G06N3/08G06V40/1347G06V40/193

Abstract

Claims (54)

1. A process, comprising:

receiving a biometric representation in a first format;

determining a dimension parameter based on the biometric representation, wherein the dimension parameter does not exceed a dimension of the biometric representation;

generating a common biometric representation in a second format by applying a feature-to-feature mapping function to the biometric representation, wherein the feature-to-feature mapping function is based on the dimension parameter and comprises a deep neural network;

training the deep neural network on a training dataset comprising a plurality of mated and non-mated biometric images associated with a plurality of human subjects; and

applying a lossy transformation to the common biometric representation to generate a token.

2. The process of claim 1 , wherein a vector dimension of the token is less than the dimension parameter.

3. A process, comprising:

receiving a biometric representation in a first format;

determining a dimension parameter based on the biometric representation, wherein the dimension parameter does not exceed a dimension of the biometric representation;

generating a common biometric representation in a second format by applying a feature-to-feature mapping function to the biometric representation, wherein the feature-to-feature mapping function is based on the dimension parameter and comprises a deep neural network;

generating a training dataset comprising a plurality of mated and non-mated synthetic biometric images, wherein the training dataset excludes biometric data associated with real human subjects; and

training the deep neural network on the training dataset.

4. The process of claim 3 , wherein sets of mated biometric images of the training dataset each comprise at least one biometric image associated with an optimal condition and at least one biometric image associated with a non-optimal condition.

5. The process of claim 4 , wherein the non-optimal condition is an underlit lighting condition.

6. The process of claim 4 , wherein the non-optimal condition is an adverse backlight condition.

7. The process of claim 4 , wherein the non-optimal condition is an overlit lighting condition.

8. The process of claim 4 , wherein the non-optimal condition is a rotation condition.

9. The process of claim 4 , wherein:

the plurality of mated and non-mated synthetic biometric images comprise facial images;

the optimal condition is a first facial expression; and

the non-optimal condition is a second facial expression different from the first facial expression.

10. The process of claim 3 , further comprising applying a lossy transformation to the common biometric representation to generate a token.

11. The process of claim 10 , wherein a vector dimension of the token is less than the dimension parameter.

12. A system, comprising:

at least one processor in communication with at least one data store;

the at least one data store comprising:

a feature-to-feature mapping function that, when applied, transforms biometric representations from a first format to a common format; and

a dimensionality reduction function that, when applied, reduces a dimension of biometric representations in the common format to a dimension parameter;

a non-transitory, machine-readable memory device comprising instructions stored thereon that, when executed by the at least one processor, cause the at least one processor to:

obtain a first biometric representation in the first format;

obtain a second biometric representation in the common format, wherein the first biometric representation is associated with a first subject and the second biometric representation is associated with a second subject;

determine the dimension parameter for the common format based on the first biometric representation and the second biometric representation, wherein the dimension parameter does not exceed a vector size of the first biometric representation or the second biometric representation;

apply the feature-to-feature mapping function to the first biometric representation to generate a first common biometric representation;

apply the dimensionality reduction function to the second biometric representation to generate a second common biometric representation, wherein the first common biometric representation and the second common biometric representation are of a second vector size equal to the dimension parameter;

compare the first common biometric representation to the second common biometric representation;

based on the comparison, determine that the first common biometric representation is within a similarity threshold of the second common biometric representation; and

transmit, to a computing device, a positive verification of a match between the first subject and the second subject.

13. The system of claim 12 , wherein:

the feature-to-feature mapping function comprises a deep neural network; and

the instructions, when executed by the at least one processor, further cause the at least one processor to train the deep neural network on a first training dataset comprising a plurality of mated and non-mated biometric representations associated with human subjects.

14. The system of claim 13 , wherein the instructions, when executed by the at least one processor, further cause the at least one processor to:

generate a second training dataset comprising a plurality of mated and non-mated synthetic biometric representations; and

train the deep neural network on the second training dataset.

15. The system of claim 14 , wherein the instructions, when executed by the at least one processor, further cause the at least one processor to:

generate a third training dataset comprising at least a portion of the first training dataset and the second training dataset; and

train the deep neural network on the third training dataset.

16. A non-transitory, computer-readable medium comprising instructions that, when executed by a computer, cause the computer to:

obtain a first common biometric representation of a first length and in a first format;

obtain a second biometric representation of a second length and in a second format, wherein the first length exceeds the second length;

apply a feature-to-feature mapping function to the second biometric representation to transform the second biometric representation into a second common biometric representation in the first format, wherein the second common biometric representation comprises a third length less than the first length and the second length; and

apply a dimensionality reduction function to the first common biometric representation to reduce the first common biometric representation from the first length to the third length.

17. The non-transitory, computer-readable medium of claim 16 , wherein the instructions, when executed by the computer, cause the computer to apply a lossy transformation to each of the first common biometric representation and the second common biometric representation to generate a first token and a second token.

18. The non-transitory, computer-readable medium of claim 17 , wherein the instructions, when executed by the computer, cause the computer to positively verify an identity of a subject associated with the second biometric representation based on a comparison between the first token and the second token.

Assignments (3)

Continuity (1)

References Cited (206)