Method and system for securely deploying an artificial intelligence model

1 . A method for securely deploying an artificial intelligence (AI) model, comprising:

attacking, by a model assessment engine, at least one substitute model using training data, wherein the at least one substitute model represents a surrogate of the AI model;

identifying, by the model assessment engine and in response to attacking the at least one substitute model, one or more deficiencies of the AI model;

receiving, by the model assessment engine, external data, wherein the external data appears outside of distribution of the training data;

determining, by the model assessment engine, whether the external data indicates a fraudulent attack based at least in part on the one or more deficiencies; and

protecting, by the model assessment engine, the AI model from the fraudulent attack by the external data based upon the said identifying, wherein said protecting comprises blocking said external data that has not been validated by the model assessment engine from being output to the AI model.

2 . The method of claim 1 , wherein said identifying includes simulating one or more attacks on the AI model via simulated attack data, wherein the simulating includes simulating a plurality of parallel attacks to identify the one or more deficiencies of the AI model based upon combinatorial algorithms.

3 . The method of claim 2 , further comprising:

filtering or changing incoming data fed into the at least one substitute model;

labeling the incoming data; and

training the at least one substitute model using the labeled incoming data.

4 . The method of claim 1 , wherein:

said identifying includes:

establishing one or more baselines of model performance and training data profiles; and

comparing, based upon the establishing, a performance of the AI model and one or more alternative models to select a model from the AI model and the one or more alternative models for production deployment; and

said protecting includes blocking one or more users associated with the external data based upon behavioral patterns of queries of the one or more users; or

said protecting includes using a max entropy method for detecting attacks on the AI model, the max entropy method including creating multiple detectors and aggregating responses thereof.

5 . The method of claim 1 , wherein the external data includes tabular data associated with user transaction history, account history, account data, user data, transaction information, or a combination thereof.

6 . The method of claim 5 , wherein:

said identifying is based upon sample input data and sample output data, the sample input data including one or more data points and the sample output data including labels respectively applied to the one or more data points based upon the AI model, each of the labels including a decision of whether a data point is fraud or non-fraud; or

said identifying is based upon submodular sampling to explore a state space.

7 . The method of claim 5 , wherein said protecting includes generating an alert level for indicating whether the external data include one or more attacks that can deceive the AI model based upon said identifying.

8 . The method of claim 5 , further comprising generating a probability of the external data being a fraud, wherein said generating the probability includes using one or more machine learning (ML) frameworks for binary classification, the one or more ML frameworks including boosted trees, neural networks, or a combination thereof.

9 . The method of claim 5 , wherein said protecting includes:

receiving the external data from a data feed associated with a transaction, the external data including the tabular data that define a risk attribute of an entity;

converting the tabular data into an input format;

identifying manipulation of one or more elements in the tabular data to create a false positive result; and

rejecting the tabular data as a manipulated data set.

10 . The method of claim 5 , wherein said protecting includes:

receiving the external data including the tabular data;

detecting a manipulation of the tabular data to identify a false positive result;

determining that the false positive result would occur in the AI model; and

alerting the AI model to flag the tabular data for human review.

11 . A system for securely deploying an artificial intelligence (AI) model, comprising:

at least one hardware processor; and

a non-transitory computer-readable medium storing instructions that, when executed by the at least one hardware processor, instruct the at least one hardware processor to perform operations including:

attacking, by a model assessment engine, at least one substitute model using training data, wherein the at least one substitute model represents a surrogate of the AI model;

identifying, by the model assessment engine and in response to attacking the at least one substitute model, one or more deficiencies of the AI model:

receiving, by the model assessment engine, external data, wherein the external data appears outside of distribution of the training data;

determining, by the model assessment engine, whether the external data indicates a fraudulent attack based at least in part on the one or more deficiencies; and

protecting, by the model assessment engine, the AI model from the fraudulent attack by the external data based upon the said identifying, wherein said protecting comprises blocking said external data that has not been validated by the model assessment engine from being output to the AI model.

12 . The system of claim 11 , wherein said identifying includes simulating one or more attacks on the AI model via simulated attack data, wherein the simulating includes simulating a plurality of parallel attacks to identify the one or more deficiencies of the AI model based upon combinatorial algorithms.

13 . The system of claim 12 , further comprising:

filtering or changing incoming data fed into the at least one substitute model;

labeling the incoming data; and

training the at least one substitute model using the labeled incoming data.

14 . The system of claim 11 , wherein:

said identifying includes:

establishing one or more baselines of model performance and training data profiles; and

comparing, based upon the establishing, a performance of the AI model and one or more alternative models to select a model from the AI model and the one or more alternative models for production deployment; and

said protecting includes blocking one or more users associated with the external data based upon behavioral patterns of queries of the one or more users; or

said protecting includes using a max entropy method for detecting attacks on the AI model, the max entropy method including creating multiple detectors and aggregating responses thereof.

15 . The system of claim 11 , wherein the external data includes tabular data associated with user transaction history, account history, account data, user data, transaction information, or a combination thereof.

16 . The system of claim 15 , wherein:

said identifying is based upon sample input data and sample output data, the sample input data including one or more data points and the sample output data including labels respectively applied to the one or more data points based upon the AI model, each of the labels including a decision of whether a data point is fraud or non-fraud; or

said identifying is based upon submodular sampling to explore a state space.

17 . The method of claim 5 , wherein said protecting includes generating an alert level for indicating whether the external data include one or more attacks that can deceive the AI model based upon said identifying.

18 . The system of claim 15 , further comprising generating a probability of the external data being a fraud, wherein said generating the probability includes using one or more machine learning (ML) frameworks for binary classification, the one or more ML frameworks including boosted trees, neural networks, or a combination thereof.

19 . The system of claim 15 , wherein said protecting includes:

receiving the external data from a data feed associated with a transaction, the external data including the tabular data that define a risk attribute of an entity;

converting the tabular data into an input format;

identifying manipulation of one or more elements in the tabular data to create a false positive result; and

rejecting the tabular data as a manipulated data set.

20 . The system of claim 15 , wherein said protecting includes:

receiving the external data including the tabular data;

detecting a manipulation of the tabular data to identify a false positive result;

determining that the false positive result would occur in the AI model; and

alerting the AI model to flag the tabular data for human review.