Granted Patent B2

US 12,321,498 · App. 18/147,769 · Granted Jun 3, 2025

Inventors: Ofir Ezrielev (Be'er Sheva, IL); Amihai Savir (Newton, MA); Tomer Kushnir (Omer, IL)

Assignee: Dell Products L.P.

G06F21/64G06F21/57

Abstract

Claims (75)

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

making an identification, by a data processing system that is hosting an instance of the AI model and using the instance of the AI model to provide computer implemented services, that new training data for an instance of the AI model hosted on the data processing system is available to the data processing system; and

based on the identification, by the data processing system:

obtaining an anomaly level classification for the new training data;

making a determination regarding whether the anomaly level classification exceeds an anomaly level threshold;

in a first instance of the determination where the anomaly level classification exceeds the anomaly level threshold, the method further comprises and by the data processing system:

treating the new training data as comprising poisoned training data; and

using the instance of the AI model hosted on the data processing system as is without any retraining and/or repairing of the instance of the AI model hosted on the data processing system to provide the computer implemented services; and

in a second instance of the determination where the anomaly level classification does not exceed the anomaly level threshold, the method further comprises and by the data processing system:

performing learning using the new training data to obtain a new instance of the instance of the AI model hosted on the data processing system; and

using the new instance of the instance of the AI model hosted on the data processing system to provide the computer implemented services.

2. The method of claim 1 , wherein the anomaly level classification is based on a corpus of training data used to obtain the instance of the AI model hosted on the data processing system where a previous instance of the instance of the AI model hosted on the data processing system was trained using the corpus of the training data to obtain the instance of the AI model hosted on the data processing system and the corpus of the training data excluding the new training data.

3. The method of claim 2 , wherein the corpus of the training data is unlabeled.

4. The method of claim 2 , further comprising, after the new training data is treated as comprising the poisoned training data:

performing a similarity analysis between a portion of the new training data and the corpus of the training data to identify a portion of the corpus of the training data;

treating the portion of the corpus of the training data as potentially comprising second poisoned training data; and

instead of using the instance of the AI model hosted on the data processing system as is to provide the computer implemented service, using a previous instance of the instance of the AI model hosted on the data processing system that is free from influence of the second poisoned training data to provide the computer implemented services.

5. The method of claim 4 , further comprising:

determining that second new training data for the instance of the AI model hosted on the data processing system is available;

using classification metrics obtained during the similarity analysis to identify a portion of the second new training data; and

treating the portion of the second new training data as comprising third poisoned training data.

6. The method of claim 5 , wherein treating the new training data as comprising the poisoned training data further comprises:

excluding the new training data from being integrated into the corpus of the training data; and

submitting the new training data for analysis by a subject matter expert to confirm whether the new training data comprises the poisoned training data.

7. The method of claim 1 , wherein the new instance of the instance of the AI model hosted on the data processing system is based the AI model.

8. The method of claim 7 , wherein the new training data is unlabeled.

9. The method of claim 1 , further comprising and by the data processing system prior to making the identification that the new training data for the instance of the AI model hosted on the data processing system is available to the data processing system:

obtaining one or more snapshots of the instance of the AI model hosted on the data processing system, wherein the one or more snapshots are used to obtain an anomaly detection model that is used to obtain the anomaly level classification for the new training data.

10. The method of claim 1 , further comprises and by the data processing system:

obtaining one or more snapshots of the instance of the AI model hosted on the data processing system from a storage of the data processing system;

using the one or more snapshots to obtain a corpus of training data used to obtain the instance of the AI model hosted on the data processing system where a previous instance of the instance of the AI model hosted on the data processing system was trained using the corpus of the training data to obtain the instance of the AI model hosted on the data processing system; and

using the corpus of the training data and the new training data to obtain the anomaly level classification for the new training data.

11. A non-transitory machine-readable medium having instructions stored therein, which when executed by a processor of a data processing system, cause the processor to perform operations for managing an artificial intelligence (AI) model, the operations comprising:

making an identification that new training data for an instance of the AI model is available to the data processing system, the instance of the AI model being hosted on the data processing system and used by the data processing system to provide computer implemented services; and

based on the identification:

obtaining an anomaly level classification for the new training data;

making a determination regarding whether the anomaly level classification exceeds an anomaly level threshold;

in a first instance of the determination where the anomaly level classification exceeds the anomaly level threshold, the operations further comprise:

treating the new training data as comprising poisoned training data; and

using the instance of the AI model hosted on the data processing system as is without any retraining and/or repairing of the instance of the AI model hosted on the data processing system to provide the computer implemented services; and

in a second instance of the determination where the anomaly level classification does not exceed the anomaly level threshold, the method further comprises:

performing learning using the new training data to obtain a new instance of the instance of the AI model hosted on the data processing system; and

using the new instance of the instance of the AI model hosted on the data processing system to provide the computer implemented services.

12. The non-transitory machine-readable medium of claim 11 , wherein the anomaly level classification is based on a corpus of training data used to obtain the instance of the AI model hosted on the data processing system model where a previous instance of the instance of the AI model hosted on the data processing system was trained using the corpus of the training data to obtain the instance of the AI model hosted on the data processing system and the corpus of the training data excluding the new training data.

13. The non-transitory machine-readable medium of claim 12 , wherein the corpus of the training data is unlabeled.

14. The non-transitory machine-readable medium of claim 12 , wherein the operations further comprise, after the new training data is treated as comprising the poisoned training data:

performing a similarity analysis between a portion of the new training data and the corpus of the training data to identify a portion of the corpus of the training data; and

treating the portion of the corpus of the training data as potentially comprising second poisoned training data; and

instead of using the instance of the AI model hosted on the data processing system as is to provide the computer implemented service, using a previous instance of the instance of the AI model hosted on the data processing system that is free from influence of the second poisoned training data to provide the computer implemented services.

15. The non-transitory machine-readable medium of claim 14 , wherein the operations further comprise:

determining that second new training data for the instance of the AI model hosted on the data processing system is available;

using classification metrics obtained during the similarity analysis to identify a portion of the second new training data; and

treating the portion of the second new training data as comprising third poisoned training data.

16. The non-transitory machine-readable medium of claim 15 , wherein treating the new training data as comprising the poisoned training data further comprises:

excluding the new training data from being integrated into the corpus of the training data; and

submitting the new training data for analysis by a subject matter expert to confirm whether the new training data comprises the poisoned training data.

17. A data processing system, comprising:

a processor; and

a memory coupled to the processor to store instructions, which when executed by the processor, cause the processor to perform operations for managing an instance of an artificial intelligence (AI) model that is hosted on the data processing system and used by the data processing system to provide computer implemented services, the operations comprising:

making an identification that new training data for the instance of the AI model hosted on the data processing system is available to the data processing system; and

based on the identification:

obtaining an anomaly level classification for the new training data;

making a determination regarding whether the anomaly level classification exceeds an anomaly level threshold;

in a first instance of the determination where the anomaly level classification exceeds the anomaly level threshold, the operations further comprise:

treating the new training data as comprising poisoned training data;

using the instance of the AI model hosted on the data processing system as is without any retraining and/or repairing of the instance of the AI model hosted on the data processing system to provide the computer implemented services; and

in a second instance of the determination where the anomaly level classification does not exceed the anomaly level threshold, the comprises operations further comprise:

performing learning using the new training data to obtain a new instance of the instance of the AI model hosted on the data processing system; and

using the new instance of the instance of the AI model hosted on the data processing system to provide the computer implemented services.

18. The data processing system of claim 17 , wherein the anomaly level classification is based on a corpus of training data used to obtain the instance of the AI model hosted on the data processing system where a previous instance of the instance of the AI model hosted on the data processing system was trained using the corpus of the training data to obtain the instance of the AI model hosted on the data processing system and the corpus of the training data excluding the new training data.

19. The data processing system of claim 18 , wherein the corpus of the training data is unlabeled.

20. The data processing system of claim 18 , wherein the operations further comprise, after the new training data is treated as comprising the poisoned training data:

performing a similarity analysis between a portion of the new training data and the corpus of the training data to identify a portion of the corpus of the training data; and

treating the portion of the corpus of the training data as potentially comprising second poisoned training data; and

instead of using the instance of the AI model hosted on the data processing system as is to provide the computer implemented service, using a previous instance of the instance of the AI model hosted on the data processing system model that is free from influence of the second poisoned training data to provide the computer implemented services.

Assignments (1)

Continuity (1)

References Cited (52)