US-12626255-B1 - Face grouping for fraud reduction

US12626255B1US 12626255 B1US12626255 B1US 12626255B1US-12626255-B1

Abstract

Using face grouping for fraud reduction is described. A server(s) may receive, in association with a request to access the service, image data representing a face of a user and additional data, determine, based at least in part on the image data and using a trained machine learning model(s), a representation of the face, and determine one or more representations of faces associated with the representation. The server(s) can further determine first information based at least in part on the additional data, determine second information associated with the one or more representations, determine one or more differences between the first information and the second information, and determine whether to accept or deny the request based at least in part on the determining of the one or more differences.

Inventors

Mitchell Jablonski
Erin Gluck
Cole Clifford
Hyunjin Choi
Kyle DeFreitas
David Puldon
Ryan Fechte
Danielle Fiudo
Sachin Rana
Aditya Joshi

Assignees

BLOCK, INC.

Dates

Publication Date: 20260512
Application Date: 20221108

Claims (15)

1 . A system comprising: one or more processors; and memory coupled to the one or more processors, with computer-executable instructions stored thereon that, when executed by the one or more processors, cause the one or more processors to perform operations comprising: determining an accuracy metric associated with a first trained machine learning model that is trained to generate representations of input faces wherein individual representations have N dimensions in a latent space; based at least in part on the accuracy metric, retraining the first trained machine learning model to obtain a second trained machine learning model that is trained to generate different dimensional representations of input faces than the first trained machine learning model wherein the different dimensional representations have fewer or more dimensions than N, wherein in response to the accuracy metric satisfying a first threshold that indicates the first trained machine learning model generates the representations of the input faces above a first accuracy, the second trained machine learning model is trained to generate P-dimensional representations of input faces, P being less than N; and wherein in response to the accuracy metric satisfying a second threshold that indicates the first trained machine learning model generates the representations of the input faces below a second accuracy, the second trained machine learning model is trained to generate Q-dimensional representations of input faces, N being less than Q; receiving, in association with a request to access a service and after the retraining, image data representing a face of a user and additional data; determining, based at least in part on the image data and using the second trained machine learning model, a representation of the face of the user, wherein the representation comprises an embedding that represents the face of the user; determining one or more representations of faces associated with the representation based on performing a search of a plurality of stored representations of faces with the representation of the face of the user as a search query for the search, wherein the one or more representations of faces comprise one or more embeddings that each represent a respective face of the faces, the one or more representations of faces being associated with one or more user accounts and previously obtained by the service; determining first information associated with the user and based at least in part on the additional data; determining second information associated with the one or more representations, the second information being associated with the one or more user accounts and previously obtained by the service; determining one or more differences between the first information and the second information; and determining whether to accept or deny the request based at least in part on of the one or more differences.
2 . The system of claim 1 , the operations further comprising: classifying the request as a fraudulent request based at least in part on the one or more differences, wherein the determining whether to accept or deny the request is based at least in part on the classifying of the request as the fraudulent request.
3 . The system of claim 1 , wherein: the image data and the additional data are received from an application executing on an electronic device, the application associated with the service; and the additional data comprises device data associated with the electronic device.
4 . The system of claim 1 , wherein: the second information includes one or more tokens stored in association with the one or more user accounts, wherein the one or more tokens indicate a number of previous requests by the one or more user accounts or by a device associated with the one or more user accounts to access the service, wherein determining whether to accept or deny the request is based at least in part on the one or more differences and is based at least in part on the one or more tokens.
5 . The system of claim 1 , wherein performing the search of the plurality of stored representations of faces includes performing, based at least in part on the representation, an approximate nearest neighbor search of the plurality of stored representations of faces stored in a data store, and wherein the determining of the one or more representations is based at least in part on the approximate nearest neighbor search.
6 . The system of claim 5 , wherein the representation is a first representation and the face of the user is a first face, the operations further comprising: determining, based at least in part on the approximate nearest neighbor search, a superset of face representations of the plurality of stored representations of faces, the superset including at least a second representation of a second face and a third representation of a third face; calculating a first distance between the first representation and the second representation; calculating a second distance between the first representation and the third representation; determining that the first distance fails to satisfy a threshold distance; and determining that the second distance satisfies the threshold distance; wherein the one or more representations: include the second representation based at least in part on the first distance failing to satisfy the threshold distance; and exclude the third representation based at least in part on the second distance satisfying the threshold distance.
7 . A computer-implemented method comprising: determining, by a server computing device, an accuracy metric associated with a first trained machine learning model that is trained to generate representations of input faces wherein individual representations have N dimensions in a latent space; based at least in part on the accuracy metric, retraining, by the server computing device, the first trained machine learning model to obtain a second trained machine learning model that is trained to generate different dimensional representations of input faces than the first trained machine learning model wherein the different dimensional representations have fewer or more dimensions than N, wherein in response to the accuracy metric satisfying a first threshold that indicates the first trained machine learning model generates the representations of the input faces above a first accuracy, the second trained machine learning model is trained to generate P-dimensional representations of input faces, P being less than N; and wherein in response to the accuracy metric satisfying a second threshold that indicates the first trained machine learning model generates the representations of the input faces below a second accuracy, the second trained machine learning model is trained to generate Q-dimensional representations of input faces, N being less than Q; receiving, by the server computing device, in association with a request to access a service and after the retraining, image data representing a face of a user and additional data associated with the user; determining, by the server computing device, and based at least in part on the image data and using the second trained machine learning model, a representation of the face of the user, wherein the representation comprises an embedding that represents the face of the user; determining, by the server computing device, one or more representations of faces associated with the representation, based on performing a search of a plurality of stored representations of faces based on distances between points in a latent space of the representation and corresponding points in a respective latent space of the each of the stored representations of faces, wherein the one or more representations of faces comprise one or more embeddings that each represent a respective face of the faces, the one or more representations of faces being associated with one or more user accounts and previously obtained by the service; determining, by the server computing device, first information associated with the user and based at least in part on the additional data; determining, by the server computing device, second information associated with the one or more representations, the second information being associated with the one or more user accounts and previously obtained by the service; determining, by the server computing device, one or more differences between the first information and the second information; and determining, by the server computing device, whether to accept or deny the request based at least in part on the one or more differences.
8 . The computer-implemented method of claim 7 , wherein: the first information comprises at least one of a first name or a first date of birth (DOB); and the second information comprises at least one of a second name or a second DOB.
9 . The computer-implemented method of claim 7 , wherein the request to access the service is associated with an identity verification (IDV) attempt.
10 . The computer-implemented method of claim 7 , wherein the image data corresponds to an image of the face of the user captured via a camera of an electronic device while the electronic device is executing an application associated with the service.
11 . The computer-implemented method of claim 7 , wherein: the embedding comprises a numerical representation of the face of the user; and the one or more embeddings comprise one or more numerical representations of a respective face of the faces.
12 . The computer-implemented method of claim 7 , wherein performing the search of the plurality of stored representations of faces includes performing, by the server computing device, and based at least in part on the representation, an approximate nearest neighbor search of the plurality of stored representations of faces stored in a data store, and wherein the determining of the one or more representations is based at least in part on the approximate nearest neighbor search.
13 . The computer-implemented method of claim 7 , wherein the representation is a first representation and the face of the user is a first face, the computer-implemented method further comprising: determining, by the server computing device, and based at least in part on an approximate nearest neighbor search, a superset of face representations of the one or more representations of faces, the superset including at least a second representation of a second face and a third representation of a third face; calculating, by the server computing device, a first distance between the first representation and the second representation; calculating, by the server computing device, a second distance between the first representation and the third representation; determining, by the server computing device, that the first distance fails to satisfy a threshold distance; and determining, by the server computing device, that the second distance satisfies the threshold distance; wherein the one or more representations: include the second representation based at least in part on the first distance failing to satisfy the threshold distance; and exclude the third representation based at least in part on the second distance satisfying the threshold distance.
14 . The system of claim 1 , wherein the search the search of the plurality of stored representations of faces is based on distances between points in a latent space of the representation and corresponding points in a respective latent space of each of the plurality of stored representations of faces, and the operations further comprise: selecting the one or more representations of faces from the plurality of stored representations of faces based on a threshold distance and based on the search.
15 . The system of claim 1 , wherein the operations further comprise: comparing the one or more differences to a threshold, wherein determining whether to accept or deny the request is based at least in part on the determining of the one or more differences and the comparing the one or more differences to the threshold.

Description

TECHNICAL FIELD Applications, which are downloadable and executable on user devices, enable users to interact with other users. Such applications are provided by service providers and utilize one or more network connections to transmit data among and between user devices to facilitate such interactions. Many applications have access to or are integrated with cameras or other sensor devices that enable image capture or the like. BRIEF DESCRIPTION OF THE DRAWINGS Features of the present disclosure, its nature and various advantages, will be more apparent upon consideration of the following detailed description, taken in conjunction with the accompanying drawings. FIG. 1 is an example environment for using face grouping for fraud reduction, according to an implementation of the present subject matter. FIG. 2 is an example diagram illustrating a face grouping technique, according to an implementation of the present subject matter. FIG. 3 is an example diagram illustrating a technique for determining differences between first information associated with a requesting user who submitted a face image and second information associated with existing user accounts that were identified using a face grouping technique, according to an implementation of the present subject matter. FIG. 4 is an example process for using face grouping for fraud reduction, according to an implementation of the present subject matter. FIG. 5 is an example process for grouping a face representation of a requesting user with similar face representations, according to an implementation of the present subject matter. FIG. 6 is an example process for retraining a machine learning model(s) to reduce computing resource consumption and/or to reduce latency without compromising face grouping accuracy, according to an implementation of the present subject matter. FIG. 7 is an example environment for performing techniques described herein. FIG. 8 is an example environment for performing techniques described herein. FIG. 9 is an example data store used for performing techniques described herein. FIG. 10 is an example environment for performing techniques described herein. FIG. 11 is an example block diagram illustrating a system for performing techniques described herein. In the figures, the left-most digit(s) of a reference number identifies the figure in which the reference number first appears. The use of the same reference numbers in different figures indicates similar or identical items or features. The drawings are not to scale. DETAILED DESCRIPTION Described herein are, among other things, techniques, devices, and systems for using face grouping for fraud reduction. In an example, a computing platform may be used to implement a service, such as a payment service. Data may be received by the computing platform (hereinafter, “service computing platform”) in association with a request to access the service. For example, during an onboarding process, as part of an identity verification (IDV) procedure for creating a new user account with the service, a user may capture an image, video, or other content representative of their face (e.g., the user may take a selfie, may have another user capture an image, video, etc. of the user, etc.) using a camera of an electronic device, and the corresponding image data may be received by the service computing platform in association with a request to create the new user account. In some examples, the data received by the service computing platform in association with the request may include additional data, such as additional image data representing another form of identification, such as a driver license, a school identification (id) card, a membership card, etc. (e.g., an image of the front of the driver license, id card, membership card, etc. and/or an image of the back of the driver license, id card, membership card, etc.). The driver license, for example, may include another face image, a name, a date of birth (DOB), an address, a machine-readable code (e.g., a barcode), and/or other information, data, symbols, indicia, or the like. Even without receiving any additional or alternative data, the service computing platform can use face grouping to verify the identity of the user (e.g., by determining, based at least in part on the image of the face of the user, that the user is who he/she says he/she is). In an example, the disclosed face grouping techniques may be used to determine whether the service computing platform has “seen” the requesting user's face in the past (e.g., whether the service computing platform has previously received face images that are similar to the requesting user's face), and, if so, this may be indicative of a repeat attempt to access the service. In some cases, detecting that the requesting user's face has been “seen” by the service computing platform in the past may be indicative of the user attempting to engage in fraudulent (or non-compliant) behavior (e.g., by creating an illeg