# US 12,002,474 B2 · Captcha automated assistant

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

- Number: US 12,002,474 B2
- Document ID: US12002474B2
- Status: Granted
- Kind: B2
- Application number: 17/737,264
- Filed: 2022-05-05
- Published: 2022-08-18
- Granted: 2024-06-04
- Earliest claimed: 2022-05-05
- Assignees: Google Inc.
- Applicants: Google LLC
- Inventors: Victor Carbune; Pedro Gonnet Anders
- Web page: https://patentgrep.com/patent/US12002474B2

## CPC

- G10L 17/04: Training, enrolment or model building
- G06F 21/31: User authentication
- G06F 40/30: Semantic analysis
- G10L 17/22: Interactive procedures; Man-machine interfaces
- G10L 17/26: Recognition of special voice characteristics, e.g. for use in lie detectors; Recognition of animal voices
- G10L 25/51: for comparison or discrimination
- H04L 63/00: {Network architectures or network communication protocols for network security (cryptographic mechanisms or cryptographic arrangements for secret or secure communication H04L9/00; network architectures or network communication protocols for wireless network security H04W12/00; security arrangements for protecting computers or computer systems against unauthorised activity G06F21/00)}
- H04W 12/00: Security arrangements; Authentication; Protecting privacy or anonymity
- G06F 2221/2133: Verifying human interaction, e.g., Captcha
- 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)
- H04L: TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION (arrangements common to telegraphic and telephonic communication H04M)
- H04W: WIRELESS COMMUNICATION NETWORKS (broadcast communication H04H; communication systems using wireless links for non-selective communication, e.g. wireless extensions H04M1/72)

## Abstract

Implementing and applying an adaptive and self-training CAPTCHA (“Completely Automated Public Turing test to tell Computers and Humans Apart”) assistant that distinguishes between a computer-generated communication (e.g., speech and/or typed) and communication that originates from a human. The CAPTCHA assistant utilizes a generative adversarial network that is self-training and includes a generator to generate synthetic answers and a discriminator to distinguish between human answers and synthetic answers. The trained discriminator is applied to potentially malicious remote entities, which are provided challenge phrases. Answers from the remote entities are provided to the discriminator to predict whether the answer originated from a human or was computer-generated.

## Claims

### Claim 1

A method for using a completely automated public Turing test (“CAPTCHA”) to classify a remote entity as malicious, the method implemented using one or more processors, comprising:
generating a textual challenge phrase using a challenge generator machine learning model, wherein the textual challenge phrase includes a challenge configured to elicit a response from a natural language dialog software agent;
transmitting the textual challenge phrase to the remote entity over one or more computing networks;
receiving an answer from the remote entity over the one or more computing networks;
applying data indicative of the answer across a trained discriminator machine learning model to generate output, wherein the discriminator machine learning model is trained in tandem with the challenge generator machine learning model; and
classifying the remote entity as malicious based on the output.

### Claim 2 (depends on 1)

The method of claim 1, wherein the textual challenge is generated by applying a ground truth textual request as input across the challenge generator machine learning model.

### Claim 3 (depends on 2)

The method of claim 2, wherein the applying further includes applying a random vector as input across the challenge generator machine learning model.

### Claim 4 (depends on 3)

The method of claim 3, wherein the random vector comprises one or more random words, phonemes, or substitutions for words in the ground truth textual request.

### Claim 5 (depends on 1)

The method of claim 1, further comprising preventing subsequent dialog with the remote entity based on the classifying.

### Claim 6 (depends on 1)

The method of claim 1, wherein the applying further comprising applying data indicative of the textual challenge phrase as input across the trained discriminator machine learning model, and the output is indicative of whether the answer is a valid response to the textual challenge phrase.

### Claim 7

A method implemented by one or more processors for using a completely automated public Turing test (“CAPTCHA”) to determine whether an incoming communication from a remote entity is computer-generated or human- generated, the method comprising:
generating a textual challenge phrase using a challenge generator machine learning model that is trained in tandem with an answer discriminator machine learning model to generate textual challenges phrases that are configured to elicit a response indicative of whether the remote entity is human or synthetic;
providing the textual challenge phrase to the remote entity;
receiving, in response to providing the textual challenge phrase, an answer from the remote entity;
assigning data indicative of the answer as input across the answer discriminator machine learning model to generate a CAPTCHA prediction, wherein the answer discriminator machine learning model is trained to generate the CAPTCHA prediction by:
providing the answer discriminator machine learning model with synthetic answers and human answers, wherein the synthetic answers include labels that identify them as synthetic and the human answers include labels that identify them as human-generated,
comparing CAPTCHA predictions output by the answer discriminator machine learning model with the labels, and
training the answer discriminator machine learning model based on the comparing;
determining, based on output generated by the answer discriminator machine learning model, a CAPTCHA prediction of whether the answer was computer-generated or human-generated;
classifying the remote entity as synthetic based on the CAPTCHA prediction; and
based on the classification of the remote entity as synthetic, preventing subsequent dialog with the remote entity.

### Claim 8 (depends on 7)

The method of claim 7, further comprising:
receiving a textual request, wherein the textual request is based on a user action of a user;
applying the textual request as input across the challenge generator machine learning model; and
receiving the textual challenge phrase from the challenge generator machine learning model, wherein the textual challenge phrase is based on the textual request.

### Claim 9 (depends on 7)

The method of claim 7, further comprising:
receiving a user action, wherein the user action includes a request to interact with a human;
generating a textual request based on the user action; and
providing the textual request as the textual challenge phrase for presentation to the human.

### Claim 10

A system for using a completely automated public Turing test (“CAPTCHA”) to classify a remote entity as malicious, the system including one or more processors to:
generate a textual challenge phrase using a challenge generator machine learning model, wherein the textual challenge phrase includes a challenge configured to elicit a response from a natural language dialog software agent;
transmit the textual challenge phrase to the remote entity over one or more computing networks;
receive an answer from the remote entity over the one or more computing networks;
apply data indicative of the answer across a trained discriminator machine learning model to generate output, wherein the discriminator machine learning model is trained in tandem with the challenge generator machine learning model; and
classify the remote entity as malicious based on the output.

### Claim 11 (depends on 10)

The system of claim 10, wherein the textual challenge is generated by applying a ground truth textual request as input across the challenge generator machine learning model.

### Claim 12 (depends on 11)

The system of claim 11, wherein the one or more processors are to apply a random vector as input across the challenge generator machine learning model.

### Claim 13 (depends on 12)

The system of claim 12, wherein the random vector comprises one or more random words, phonemes, or substitutions for words in the ground truth textual request.

### Claim 14 (depends on 10)

The system of claim 10, wherein the one or more processors are to prevent subsequent dialog with the remote entity based on the classifying.

### Claim 15 (depends on 10)

The system of claim 10, wherein the one or more processors are to apply data indicative of the textual challenge phrase as input across the trained discriminator machine learning model, and the output is indicative of whether the answer is a valid response to the textual challenge phrase.