Method and system for artificial intelligence model training using a watermark-enabled kernel for a data processing accelerator

1. A computer-implemented method performed by a data processing (DP) accelerator, the method comprising:

receiving, at the DP accelerator, first data representing a set of training data from a host, wherein the set of training data is used to train an artificial intelligence (AI) model to perform inferences;

receiving, at the DP accelerator, a watermark kernel from the host;

executing the watermark kernel within the DP accelerator on the AI model, wherein the watermark kernel, when executed, is configured to:

generate a new watermark by inheriting an existing watermark from an existing AI model,

train the AI model using the set of training data for the AI model to perform the inferences, and

implant the new watermark within the AI model during training of the AI model; and

transmitting second data representing the trained AI model having the new watermark implanted therein to the host.

2. The method of claim 1 , further comprising:

receiving a pre-trained AI model; and

wherein the training is performed for the pre-trained AI model.

3. The method of claim 1 , further comprising receiving the existing watermark from the host.

4. The method of claim 1 , wherein implanting the new watermark in the AI model comprises embedding the watermark in one or more nodes of the AI model.

5. The method of claim 4 , wherein the new watermark is stored in one or more weight variables of the one or more nodes of the AI model.

6. The method of claim 4 , wherein the new watermark is stored in one or more bias variables corresponding to the one or more nodes of the AI model, wherein each of the one or more nodes includes a weight variable and a bias variable.

7. The method of claim 1 , wherein implanting the new watermark in the AI model comprises creating one or more additional nodes of the AI model during the training to store the new watermark.

8. The method of claim 1 , wherein the host is a computing device and the DP accelerator is a general-purpose processing unit (GPU).

9. The method of claim 1 , wherein the host and DP accelerator communicate over a link that comprises a peripheral component interconnect express (PCIe) link.

10. A data processing (DP) accelerator, comprising:

an interface to receive first data representing a set of training data and to receive a watermark kernel from a host, wherein the set of training data is used to train an artificial intelligence (AI) model to perform inferences; and

a kernel executor to execute the watermark kernel within the DP accelerator on the AI model, wherein the watermark kernel, when executed, is configured to:

generate a new watermark by inheriting an existing watermark from an existing AI model,

train the AI model using the set of training data for the AI model to perform the inferences, and

implant the new watermark within the AI model during training of the AI model,

wherein second data representing the trained AI model having the new watermark implanted therein is transmitted to the host.

11. The DP accelerator of claim 10 , wherein the existing watermark is received from the host.

12. The DP accelerator of claim 10 , wherein in implanting the new watermark in the AI model, the kernel executor is configured to embed the watermark in one or more nodes of the AI model.

13. The DP accelerator of claim 12 , wherein the new watermark is stored in one or more weight variables of the one or more nodes of the AI model.

14. The DP accelerator of claim 12 , wherein the new watermark is stored in one or more bias variables corresponding to the one or more nodes of the AI model, wherein each of the one or more nodes includes a weight variable and a bias variable.

15. The DP accelerator of claim 10 , wherein implanting the new watermark in the AI model comprises creating one or more additional nodes of the AI model during the training to store the new watermark.

16. The DP accelerator of claim 10 , wherein the host is a computing device and the DP accelerator is a general-purpose processing unit (GPU).

17. The DP accelerator of claim 10 , wherein the host and DP accelerator communicate over a link that comprises a peripheral component interconnect express (PCIe) link.

18. A non-transitory machine-readable medium having instructions stored therein, which when executed by a data processing (DP) accelerator, cause the DP accelerator to perform operations, the operations comprising:

receiving, at the DP accelerator, first data representing a set of training data from a host, wherein the set of training data is used to train an artificial intelligence (AI) model to perform inferences;

receiving, at the DP accelerator, a watermark kernel from the host;

executing the watermark kernel within the DP accelerator on the AI model, wherein the watermark kernel, when executed, is configured to:

generate a new watermark by inheriting an existing watermark from an existing AI model,

train the AI model using the set of training data for the AI model to perform the inferences, and

implant the new watermark within the AI model during training of the AI model; and

transmitting second data representing the trained AI model having the new watermark implanted therein to the host.

19. The machine-readable medium of claim 18 , wherein the operations further comprise receiving the existing watermark from the host.

20. The machine-readable medium of claim 18 , wherein implanting the new watermark in AI model comprises embedding the watermark in one or more nodes of the AI model.