Watermark unit for a data processing accelerator

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

receiving, by the DP accelerator, first data, wherein the first data comprises a set of training data from a host processor;

training, by the DP accelerator, an artificial intelligence (AI) model based on the received first data using an artificial intelligence (AI) unit of the DP accelerator;

receiving, by the DP accelerator, a request to implant a watermark into the trained AI model from the host processor, wherein the request to implant the watermark includes a watermark algorithm identifier (ID) corresponding to a watermark algorithm, wherein the watermark algorithm ID is one of a plurality of watermark algorithm IDs corresponding to a plurality of watermark algorithms stored in the DP accelerator;

generating, by the DP accelerator, the watermark using the watermark algorithm associated with the received watermark algorithm ID via a watermark unit of the DP accelerator;

implanting, by the DP accelerator, the watermark within the trained AI model based on the watermark algorithm ID using the watermark unit; and

transmitting, by the DP accelerator, second data, wherein the second data comprises the trained AI model implanted with the watermark to the host processor.

2. The method of claim 1 , further comprising:

receiving, by the DP accelerator, a pre-trained AI model; and

wherein training the artificial intelligence (AI) model further comprises training the pre-trained AI model based on the received first data.

3. The method of claim 1 , further comprising:

extracting, by the DP accelerator, the watermark algorithm identifier (ID) from the request to implant the watermark.

4. The method of claim 1 , prior to receiving the request to implant the watermark, the method further comprising:

transmitting, by the DP acceleration, a notification to the host processor, wherein the notification comprises information that the AI model has been trained.

5. The method of claim 2 , further comprising:

receiving, by the DP accelerator, a watermark algorithm identifier (ID) from the host processor, wherein the host processor selects a watermark algorithm that is supported by the DP accelerator.

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

7. The method of claim 6 , wherein the watermark is implanted in one or more weight variables of the one or more nodes of the AI model.

8. The method of claim 6 , wherein the watermark is stored in one or more bias variables of the one or more nodes of the AI model.

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

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

one or more processors;

an artificial intelligence (AI) unit; and

a watermark unit;

wherein the one or more processors are configured to:

receive first data, wherein the first data comprises a set of training data from a host processor;

train an artificial intelligence (AI) model based on the received first data using the AI unit;

receive a request to implant a watermark in the trained AI model from the host processor, wherein the request to implant the watermark includes a watermark algorithm identifier (ID) corresponding to a watermark algorithm, wherein the watermark algorithm ID is one of a plurality of watermark algorithm IDs corresponding to a plurality of watermark algorithms stored in the DP accelerator;

generate, using the watermark unit, the watermark using the watermark algorithm associated with the received watermark algorithm ID;

implanting the watermark within the trained AI model based on the watermark algorithm ID using the watermark unit; and

transmit second data, wherein the second data comprises the trained AI model implanted with the watermark to the host processor.

11. The DP accelerator of claim 10 , wherein the one or more processors are further configured to:

receive a pre-trained AI model; and

wherein training the AI model further comprises training the pre-trained AI model based on the received first data.

12. The DP accelerator of claim 10 , wherein the one or more processors are further configured to:

extract the watermark algorithm identifier (ID) from the request to implant the watermark.

13. The DP accelerator of claim 10 , wherein, prior to receiving the request to implant the watermark, the one or more processors are further configured to:

transmit a notification to the host processor, wherein the notification comprises information that the AI model has been trained.

14. The DP accelerator of claim 11 , wherein the one or more processors are further configured to:

receive a watermark algorithm identifier (ID) from the host processor, wherein the host processor selects a watermark algorithm that is supported by the DP accelerator.

15. The DP accelerator of claim 10 , wherein implanting the watermark in the AI model further comprises embedding the watermark in one or more nodes of the AI model.

16. The DP accelerator of claim 15 , wherein the watermark is implanted in one or more weight variables of the one or more nodes of the AI model.

17. The DP accelerator of claim 15 , wherein the watermark is stored in one or more bias variables of one or more nodes of the AI model.

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

19. 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 first data, wherein the first data comprises a set of training data from a host processor;

training an artificial intelligence (AI) model based on the received first data using an artificial intelligence (AI) unit of the DP accelerator;

receiving a request to implant a watermark in the trained AI model from the host processor, wherein the request to implant the watermark includes a watermark algorithm identifier (ID) corresponding to a watermark algorithm, wherein the watermark algorithm ID is one of a plurality of watermark algorithm IDs corresponding to a plurality of watermark algorithms stored in the DP accelerator;

generating the watermark using the watermark algorithm associated with the received watermark algorithm ID via a watermark unit of the DP accelerator;

implanting the watermark within the trained AI model based on the watermark algorithm ID using the watermark unit of the DP accelerator; and

transmitting second data, wherein the second data comprises the trained AI model implanted with the watermark to the host processor.

20. The machine-readable medium of claim 19 , wherein the instructions, which when executed by the DP accelerator, further cause the DP accelerator to perform the operations comprising:

extracting a watermark algorithm identifier (ID) from the request to implant a watermark.