US 2025/0028798 A1Published application

Artificial intelligence impersonation detector

AssigneeBank of America Corporation

InventorsManu Kurian, Ana Maxim, Vinesh Young Patel, Michael Young

Life of this patent

Estimating the term…

Loading patent…

Abstract

Artificial intelligence (AI) impersonation detection using an AI model is provided. Methods may train an AI model. The training may provide the AI model with a first dataset including a communication between two or more human users, a second and third dataset including an impersonation of the communication by a public AI model and a private AI model respectively. Methods may identify a first identifier set in the first dataset, a second identifier set in the second dataset and a third identifier set in the third dataset. Methods may create a parameter range for the first, second and third identifier set. The methods may include monitoring a production communication using the AI model. The AI model may compare the production communication with the parameter ranges. Methods may identify a smallest comparison value. Methods may, based on the identified smallest comparison value, identify the origin of the production communication.

Claims20

Dependent claims

A method for detecting artificial intelligence (AI) impersonation, the method comprising:

in a pre-detection stage, training a first AI model, said training comprising:
receiving:
a first dataset, the first dataset comprising a communication, the communication between two or more human users;
a second dataset, the second dataset comprising an impersonation of the communication by a second AI model, the second AI model receiving data from a private database;
and
a third dataset, the third dataset comprising an impersonation of the communication by a third AI model, the third AI model receiving data from a public database;
comparing the first, second and third datasets;
based on the comparing, identifying:
a first set of identifiers in the first dataset;
a second set of identifiers in the second dataset;
and
a third set of identifiers in the third dataset;
creating a parameter range reference, said creating comprising:
using the first set of identifiers to determine a first parameter range;
using the second set of identifiers to determine a second parameter range;
and
using the third set of identifiers to determine a third parameter range;
in a detection stage, monitoring a production communication, said monitoring using the first AI model, said monitoring comprising:
comparing the production communication to the first parameter range, said comparing outputting a first comparison value;
comparing the production communication to the second parameter range, said comparing outputting a second comparison value;
comparing the production communication to the third parameter range, said comparing outputting a third comparison value;
identifying a smallest comparison value from the first, second, and third comparison values;
identifying a parameter range from the first, second, and third parameter ranges corresponding the identified smallest comparison value;
identifying a dataset from the first, second and third datasets corresponding to the identified parameter range;
and
based on the identified dataset, detecting that the production communication is a human communication, a private AI communication or a public AI communication.

23 more elements

An apparatus for detecting artificial intelligence (AI) impersonation, the apparatus comprising:

a pre-detection monitor, the pre-detection monitor configured to train a first AI model, the first AI model configured to:
receive:
a first dataset, the first dataset comprising a communication, the communication between two or more human users;
a second dataset, the second dataset comprising an impersonation of the communication by a second AI model, the second AI model receiving data from a private database;
and
a third dataset, the third dataset comprising an impersonation of the communication by a third AI model, the third AI model receiving data from a public database;
compare the first, second and third datasets;
based on the comparison, identify:
a first set of identifiers in the first dataset;
a second set of identifiers in the second dataset;
and
a third set of identifiers in the third dataset;
create a parameter range reference, the first AI model configured to:
use the first set of identifiers to determine a first parameter range;
use the second set of identifiers to determine a second parameter range;
and
use the third set of identifiers to determine a third parameter range;
and
a detection monitor, the detection monitor configured to monitor a production communication using the first AI model, said detection monitor configured to:
compare the production communication to the first parameter range, the first AI model configured to output a first comparison value;
compare the production communication to the second parameter range, said comparison configured to output a second comparison value;
compare the production communication to the third parameter range, said comparison configured to output a third comparison value;
identify a smallest comparison value from the first, second, and third comparison values;
identify a parameter range from the first, second, and third parameter ranges corresponding the identified smallest comparison value;
identify a dataset from the first, second and third datasets corresponding to the identified parameter range;
and
based on the identified dataset, detect that the production communication is a human communication, a private AI communication or a public AI communication.

24 more elements

Dependencies are inferred from the claim text.