Granted Patent B1

US 12,401,680 · App. 19/200,467 · Granted Aug 26, 2025

Inventors: Rahul Sharma (Fremont, CA); Narasimha ‘Murthy’ M Kunchakarra (Fremont, CA)

Assignee: U.S. BANCORP, NATIONAL ASSOCIATION

H04L63/1433G06F40/30

Abstract

Claims (82)

1. A system for dynamically adjusting on-device configurations for message type vulnerability for a text-based model in a federated network by dynamically modifying context based on detected context switches in an audio mode, the system comprising one or more processors configured by computer-readable instructions to perform operations comprising:

determining, using an on-device language model executing on a first electronic device, a first context parameter based on text input derived from audio data received via an on-device audio receiver of the first electronic device;

receiving, from a first user, a user-provided input via a text-based entry on a user interface screen of the first electronic device;

generating, with the first electronic device, a first output prediction indicating message vulnerability for a message type that, if interacted with, compromises data accessible via the first electronic device or control of the first electronic device by (1) selecting a first prediction model in lieu of alternative prediction models by referencing an associative data structure associating parameters and prediction models with the first context parameter to select the first prediction model, and (2) providing, to the first prediction model, an input comprising the user-provided input to generate the first output prediction;

in response to generating the first output prediction indicating the message vulnerability to the message type:

presenting a first test message of the message type on a display of the first electronic device to test the message vulnerability of the first electronic device during later use of the first electronic device, wherein a target interaction with the first electronic device after the presenting of the first test message on the first electronic device causes the first electronic device to send a confirmation message indicating a first vulnerability status to a server; and

transmitting a version of the first prediction model to a second electronic device in a federated learning network with the first electronic device without requiring the second electronic device to independently perform training to obtain the version of the first prediction model, wherein the second electronic device is caused to generate a second output prediction indicating the message vulnerability to the message type for the second electronic device using the first prediction model, and wherein the second electronic device presents a second test message of the message type to test device vulnerability in response to generating the second output prediction.

2. The system of claim 1 , the operation further comprising determining a result indicating whether the first context parameter is associated with a category indicating that the first context parameter is temporary, wherein generating the first output prediction comprises generating the first output prediction based on the result indicating whether the first context parameter is associated with the category indicating that the first context parameter is temporary.

3. The system of claim 1 , the operation further comprising:

determining a second context parameter based on the user-provided input;

determining a correlation parameter between the first context parameter and the second context parameter based on a count of events characterized by the first context parameter being determined and the second context parameter being determined within a same predefined time interval;

determining a result indicating whether the correlation parameter satisfies a threshold; and

constructing an association between the first context parameter and the second context parameter based on the result indicating whether the correlation parameter satisfies the threshold, wherein determining the first prediction model comprises selecting the first prediction model based on the association between the first context parameter and the second context parameter.

4. A method adjusting on-device configurations for message type vulnerability based on detected context switches, comprising:

determining, using an on-device language model executing on a first electronic device, a first context parameter based on a first input derived from first data received via a first mode of the first electronic device;

receiving a second input for a prediction model via a second mode of the first electronic device;

generating a first output prediction indicating message vulnerability for a message type by (1) determining a first prediction model in lieu of alternative prediction models based on an associative data structure associating the first context parameter with the first prediction model and (2) providing, to the first prediction model, an input comprising the second input to cause the first prediction model to generate the first output prediction; and

based on the generating of the first output prediction:

presenting a first message of the message type on a display of the first electronic device, wherein a target interaction with the first electronic device after the presenting of the first message on the first electronic device causes the first electronic device to send a second message indicating a vulnerability status to a server; and

transmitting a version of the first prediction model to a second electronic device, wherein the second electronic device is caused to generate a second output prediction indicating the message vulnerability to the message type.

5. The method of claim 4 , wherein generating the first output prediction comprises:

determining, using the on-device language model, a second context parameter based on the first input;

determining a first modality weight associated with data received via the first mode; and determining a comparison result indicating that the first modality weight associated with the first mode outweighs a second modality weight associated with the second mode, and wherein selecting the first prediction model comprises selecting the first prediction model based on the comparison result.

6. The method of claim 4 , further comprising:

detecting, via the on-device language model, a previous context parameter based on the first input;

determining a previous model mapped to the previous context parameter, wherein generating the first output prediction comprises:

determining a result indicating whether a similarity between the previous model and the first prediction model satisfies a threshold; and

generating the first output prediction based on the result.

7. The method of claim 4 , further comprising:

obtaining a set of model configuration parameters by referencing a context map based on the first context parameter and the first mode; and

determining the first prediction model comprises configuring an initial model to construct the first prediction model based on the set of model configuration parameters.

8. The method of claim 4 , wherein determining the first prediction model comprises:

retrieving a set of adapter neural networks mapped to the first context parameter; and

configuring the prediction model to use a set of adapter neural network nodes.

9. The method of claim 4 , wherein determining the first prediction model comprises performing a retraining operation on a previous model based on a set of responses to previous messages stored in the first electronic device to determine the first prediction model.

10. The method of claim 4 , further comprising:

detecting a set of stable contextual variables by determining a first set of features that remain associated with particular models or outcomes across different context parameter switches;

detecting a set of exogenous contextual variables by determining a second set of features that vary across different context parameter switches; and

training the prediction model by:

training a first subset of model layers based on the set of stable contextual variables; and

training a second subset of model layers based on the set of exogenous contextual variables.

11. The method of claim 4 , further comprising

based on the target interaction, sending a response message associated with the first message to the server, wherein the response message causes a creation of a new user account.

12. One or more non-transitory, machine-readable media storing program instructions that, when executed by one or more processors, causes the one or more processors to perform operations comprising:

determining, using a first electronic device, a first context parameter based on text input derived from audio data received via an on-device audio receiver of the first electronic device;

receiving, from a first user, a user-provided input for a prediction model via a text-based entry on a user interface screen of the first electronic device;

generating an output prediction indicating message vulnerability for a message type that, if interacted with, compromises data accessible via the first electronic device or control of the first electronic device by (1) selecting a first prediction model in lieu of alternative prediction models by referencing an associative data structure associating parameters and prediction models to select the first prediction model and (2) providing, to the first prediction model, an input comprising the user-provided input to obtain the output prediction; and

based on the generating of the output prediction,

presenting a first test message of a message type on a display of the first electronic device to test the message vulnerability of the first electronic device during later use of the first electronic device, wherein a target interaction with the first electronic device after the presenting of the first test message on the first electronic device causes the first electronic device to send a confirmation message associated with the first message to a server; and

transmitting a version of the first prediction model to a second electronic device in a federated learning network with the first electronic device without requiring the second electronic device to independently perform training to obtain the version of the first prediction model, wherein the second electronic device is caused to generate a second output prediction indicating the message vulnerability to the message type for the second electronic device using the first prediction model, and wherein the second electronic device presents a second test message of the message type to test device vulnerability in response to generating the second output prediction.

13. The one or more media of claim 12 , the operations further comprising:

detecting a candidate sequence of recurring context variables comprising the first context parameter and a subsequent second context;

detecting a plurality of instances indicating the candidate sequence of recurring context variables;

determining a schedule of context based on the plurality of instances, wherein the schedule of context variables comprises the first context parameter associated with a first time interval in the schedule of contexts; and

receiving a third input from the first user causing a selection of the first prediction model in response to the third input being provided during the first time interval.

14. The one or more media of claim 12 , the operations further comprising:

detecting a temporal pattern of shifts based on the shifts;

constructing a schedule based on the temporal pattern; and

generating a future output prediction indicating message vulnerability by:

determining that a context shift is predicted for a future period; and

pre-configure a model with a set of model parameters associated with the first context parameter before or while the future period occurs.

15. The one or more media of claim 12 , the operations further comprising determining a result indicating whether the first context parameter is associated with a category indicating that the first context parameter is temporary, wherein generating the output prediction indicating message vulnerability comprises generating the output prediction based on the result indicating whether the first context parameter is associated with the category indicating that the first context parameter is temporary.

16. The one or more media of claim 12 , the operations further comprising:

determining a second context parameter based on a third input obtained via a third mode different from the on-device audio receiver and different from the text-based entry on the user interface screen;

determining a correlation parameter between the first context parameter and the second context parameter based on a count of events characterized by the first context parameter being determined and the second context parameter being determined within a same predefined time interval;

determining a result indicating whether the correlation parameter satisfies a threshold; and

constructing an association between the first context parameter and the second context parameter based on the result indicating whether the correlation parameter satisfies the threshold, wherein determining the first prediction model comprises selecting the first prediction model based on the association between the first context parameter and the second context parameter.

17. The one or more media of claim 12 , wherein generating the output prediction indicating message vulnerability comprises:

determining a second context parameter based on the text input;

determining a first modality weight associated with data received via the on-device audio receiver; and

determining a comparison result indicating that the first modality weight associated with the on-device audio receiver outweighs a second modality weight associated with the text-based entry on the user interface screen, and wherein selecting the first prediction model comprises selecting the first prediction model based on the comparison result.

18. The one or more media of claim 12 , the operations further comprising:

detecting a previous context parameter based on the text input;

determining a previous model mapped to the previous context parameter, wherein generating the output prediction indicating message vulnerability comprises:

determining a result indicating whether a similarity between the previous model and the first prediction model satisfies a threshold; and

generating the output prediction based on the result.

19. The one or more media of claim 12 , the operations further comprising

obtaining a set of model configuration parameters by referencing a context map based on the first context parameter and the on-device audio receiver; and

determining the first prediction model comprises configuring an initial model to construct the first prediction model based on the set of model configuration parameters.

20. The one or more media of claim 12 , wherein determining the first prediction model in lieu of alternative prediction models comprises:

retrieving a set of adapter neural networks mapped to the first context parameter; and

configuring the prediction model to use a set of adapter neural network nodes.

Assignments (1)

References Cited (59)

Cited By (1)

US 12,718,012