Combination of protection measures for artificial intelligence applications against artificial intelligence attacks

1. A computer implemented method of protecting an artificial intelligence (AI) application, the method comprising:

identifying parameters of the AI application;

performing an assessment of a vulnerability of the AI application, comprising:

applying a combination of protection measures comprising two or more protection measures against at least two different attacks and at least one dataset; and

determining whether the combination of protection measures is successful in defending the AI application;

determining a target configuration of an AI model to protect the AI application based on the assessed vulnerability of the AI application;

determining a machine learning algorithm to adjust the AI model to include a combination of most computationally efficient defenses based on the target configuration; and

using the adjusted AI model to protect the AI application.

2. The computer implemented method of claim 1 , wherein, for each of the at least two different attacks, a goal of the attack is received from at least one of an administrator or user of the AI application.

3. The computer implemented method of claim 1 , wherein the goal of the AI application is automatically determined by the computer by reviewing the AI application.

4. The computer implemented method of claim 1 , wherein the goal of the AI application is received by the computer from at least one of an administrator or user of the AI application.

5. The computer implemented method of claim 1 , wherein the target configuration is based on metrics of the ML model, including at least one of: accuracy, F1 score, F2 score, false positive rate, false negative rate, false sensitivity rate, empirical robustness, Cross Lipshitz Extreme Value (CLEVER) score, Misclassification Ratio (MR),Classification Rectify Ratio (CRR), Classification Sacrifice Ratio (CSR), Classification Confidence Variance (CCV), Classification Output Stability (COS), Average Confidence of Adversarial Class (ACAC), Average Confidence of True Class (ACTC), Average Lp Distortion (ALDp), Average Structural Similarity (ASS), Perturbation Sensitivity Distance (PSD), Noise Tolerance Estimation (NTE), Robustness to Gaussian Blur (RGB), or Robustness to Image Compression (RIC).

6. The computer implemented method of claim 1 , further comprising, using a pre-optimization algorithm to optimize (i) a number of attacks and (ii) a number of defenses to be used in the combination of protection measures, based on at least one of a goal of the attack, a mindmap of the attack, and an expected time of the attack.

7. The computer implemented method of claim 1 , further comprising, using an on-demand optimization algorithm to exclude an attack, a defense, or a dataset from the combination of protection measures, when the corresponding attack, defense, or dataset does not perform according to its corresponding metric.

8. The computer implemented method of claim 1 , wherein the target configuration is based on a defense category of the AI model.

9. The computer implemented method of claim 1 , further comprising updating the assessment of the vulnerability of the AI application by tracking one or more databases for documents of protection measures or defenses, by way of one or more crawlers.

10. The computer implemented method of claim 1 , further comprising updating the assessment of the vulnerability of the AI application further by, for each document, tagging parameters of the document, including at least one of a defense configuration, goal configuration, target configuration, or metric configuration, by way of a text classification algorithm.

11. A non-transitory computer readable storage medium tangibly embodying a computer readable program code having computer readable instructions that, when executed, causes a computer device to carry out a method of protecting an artificial intelligence (AI) application, the method comprising:

identifying parameters of the AI application;

performing an assessment of a vulnerability of the AI application, comprising:

applying a combination of protection measures comprising two or more protection measures against at least two different attacks and at least one dataset; and

determining whether the combination of protection measures is successful in defending the AI application;

determining a target configuration of an AI model to protect the AI application based on the assessed vulnerability of the AI application;

using a pre-optimization algorithm to optimize a number of attacks and a number of defenses to be used in the combination of protection measures, based on at least one of a goal of the attack, a mindmap of the attack, and an expected time of the attack.

12. The computer implemented method of claim 11 , further comprising:

determining a machine learning algorithm to adjust the AI model to include a combination of most computationally efficient defenses based on the target configuration; and

using the adjusted AI model to protect the AI application.

13. The computer implemented method of claim 11 , wherein, for each of the at least two different attacks, a goal of the attack is received from at least one of an administrator or user of the AI application.

14. The computer implemented method of claim 11 , wherein one or more properties of the AI application, including at least one of an AI object, a machine learning (ML) category, an ML algorithm, and an ML activation function of the AI application are automatically determined by the computer device by reviewing the AI application.

15. The computer implemented method of claim 11 , wherein one or more properties of the AI application, including at least one of an AI object, a machine learning (ML) category, an ML algorithm, and an ML activation function of the AI application are received by the computer device from at least one of an administrator or user of the AI application.

16. The computer implemented method of claim 11 , wherein the target configuration is based on metrics of the ML model, including at least one of: accuracy, F1 score, F2 score, false positive rate, false negative rate, false sensitivity rate, empirical robustness, Cross Lipshitz Extreme Value (CLEVER) score, Misclassification Ratio (MR),Classification Rectify Ratio (CRR), Classification Sacrifice Ratio (CSR), Classification Confidence Variance (CCV), Classification Output Stability (COS), Average Confidence of Adversarial Class (ACAC), Average Confidence of True Class (ACTC), Average Lp Distortion (ALDp), Average Structural Similarity (ASS), Perturbation Sensitivity Distance (PSD), Noise Tolerance Estimation (NTE), Robustness to Gaussian Blur (RGB), or Robustness to Image Compression (RIC).

17. The computer implemented method of claim 11 , further comprising updating the assessment of the vulnerability of the AI application by tracking one or more databases for documents of protection measures or defenses, by way of one or more crawlers.

18. A computing device comprising:

a processor;

a storage device coupled to the processor;

a protection engine stored in the storage device, wherein an execution of the protection engine by the processor configures the computing device to perform acts comprising:

identifying parameters of an AI application;

performing an assessment of a vulnerability of the AI application, comprising:

applying a combination of protection measures comprising two or more protection measures against at least two different attacks and at least one dataset; and

determining whether the combination of protection measures is successful in defending the AI application;

determining a target configuration of an AI model to protect the AI application based on the assessed vulnerability of the AI application;

determining an AI enhanced algorithm to adjust the AI model to include a combination of most computationally efficient defenses based on the target configuration; and

using the adjusted AI model to protect the AI application.