Systems and methods for generating and using anthropomorphic signatures to authenticate users
Inventors
SLY, Axel • SHARMA, Srivatsa Akshay • REDINGER, Brett Robert • REICH, Devin Daniel • Trooskens, Geert • LOOTUS, Meelis • Lee, Young Jin • ARREDONDO, Ricardo Lopez • KAUTZ, IV, Frederick Franklin • BHAT, Satish Srinivasan • KIRK, Scott Michael • DE BROUWER, Walter Adolf • THAKORE, Kartik
Assignees
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Publication Number
US-11321447-B2
Publication Date
2022-05-03
Expiration Date
Abstract
The technology disclosed relates to authenticating users using a plurality of non-deterministic registration biometric inputs. During registration, a plurality of non-deterministic biometric inputs are given as input to a trained machine learning model to generate sets of feature vectors. The non-deterministic biometric inputs can include a plurality of face images and a plurality of voice samples of a user. A characteristic identity vector for the user can be determined by averaging feature vectors. During authentication, a plurality of non-deterministic biometric inputs are given as input to a trained machine learning model to generate a set of authentication feature vectors. The sets of feature vectors are projected onto a surface of a hyper-sphere. The system can authenticate the user when a cosine distance between the authentication feature vector and a characteristic identity vector for the user is less than a pre-determined threshold.
Core Innovation
The invention provides a computer-implemented authentication approach that uses a plurality of non-deterministic authentication biometric inputs, including an image and a voice sample of a user, received with a request for authentication. The biometric inputs are fed to a trained machine learning model to generate a set of authentication feature vectors, which are compared to a characteristic identity vector previously registered for the user.
To perform the comparison, the set of feature vectors is projected onto a surface of a hyper-sphere. Authentication is performed when a cosine distance between the authentication feature vectors and the characteristic identity vector is less than a pre-determined threshold.
The characteristic identity vector for the user is determined by averaging feature vectors for a plurality of images and for a plurality of voice samples of the user. In further embodiments, the pre-determined threshold is determined based on variance, including variance based on projected feature vectors on a user-by-user basis, variance for user classes, or variance across users.
Alternative matching is described using distance-preserving hashing via locality-preserving hashing, and also via binning/quantization followed by hashing, as other ways to perform matching in the authentication system context alongside the hyper-sphere and cosine-distance biometric embedding approach.
Claims Coverage
The independent claims are clm-00001, clm-00007, and clm-00013. Across these claims, each includes the same core inventive scheme with multiple non-deterministic biometrics (image and voice sample), a trained machine learning model producing authentication feature vectors, hyper-sphere projection, and authentication based on cosine distance to a registered characteristic identity vector below a pre-determined threshold, where the characteristic identity vector is determined by averaging feature vectors from multiple images and multiple voice samples.
Hyper-sphere projection with cosine-distance authentication
Receiving a plurality of non-deterministic biometric inputs with a request for authentication; feeding the non-deterministic biometric inputs to a trained machine learning model and generating a set of authentication feature vectors, wherein the non-deterministic authentication biometric input includes an image and a voice sample of a user; projecting the set of feature vectors onto a surface of a hyper-sphere; authenticating the user when a cosine distance between authentication feature vectors and a characteristic identity vector previously registered for the user is less than a pre-determined threshold.
Averaged characteristic identity vector from multiple images and voice samples
The characteristic identity vector for the user is determined by averaging feature vectors for a plurality of images and for a plurality of voice samples of the user.
Threshold determination based on variance of projected feature vectors
A predetermined threshold is determined based on variance associated with projected feature vectors, including variance on a user-by-user basis, variance among projected feature vectors for classes of users, and variance among projected feature vectors across users.
Non-transitory computer readable medium implementing the authentication method
A non-transitory computer readable storage medium impressed with computer program instructions to authenticate using a plurality of non-deterministic authentication biometric inputs, implementing receiving the plurality of non-deterministic biometric inputs with a request for authentication; feeding the non-deterministic biometric inputs to a trained machine learning model and generating a set of authentication feature vectors including an image and a voice sample of a user; projecting the set of feature vectors onto a surface of a hyper-sphere; and authenticating the user when a cosine distance between authentication feature vectors and a characteristic identity vector previously registered for the user is less than a pre-determined threshold, wherein the characteristic identity vector is determined by averaging feature vectors for a plurality of images and for a plurality of voice samples of the user.
Processor-based authentication system
A system including one or more processors coupled to memory, the memory loaded with computer instructions to authenticate using a plurality of non-deterministic authentication biometric inputs, where the instructions implement receiving the plurality of non-deterministic biometric inputs with a request for authentication; feeding the non-deterministic biometric inputs to a trained machine learning model and generating a set of authentication feature vectors including an image and a voice sample of a user; projecting the set of feature vectors onto a surface of a hyper-sphere; and authenticating the user when a cosine distance between authentication feature vectors and a characteristic identity vector previously registered for the user is less than a pre-determined threshold, wherein the characteristic identity vector is determined by averaging feature vectors for a plurality of images and for a plurality of voice samples of the user.
Across independent claim coverage, the invention is characterized by authentication using multiple non-deterministic biometrics (image and voice sample), conversion into authentication feature vectors by a trained machine learning model, projection onto a hyper-sphere, and authentication by comparing cosine distance to a pre-registered characteristic identity vector below a pre-determined threshold, with the characteristic identity vector determined by averaging feature vectors from multiple images and multiple voice samples. The claim set further refines how the pre-determined threshold may be derived based on variance, including user-by-user, user-class, or across-user variance.
Stated Advantages
Documented Applications
No documented applications found
Interested in licensing this patent?