Hardware accelerator and method of operating a hardware accelerator
In accordance with a first aspect of the present disclosure, a hardware accelerator is provided, comprising: an execution unit configured to execute at least one artificial intelligence (AI) model using one or more resources of the hardware accelerator; a watermark verification unit configured to verify whether a predefined watermark is present in the AI model and to output a verification result indicative of the presence or absence of said watermark; a resource management unit configured to enable said resources if the verification result indicates that the watermark is present. Further aspects of the present disclosure relate to a corresponding method of operating a hardware accelerator, and to a computer program for carrying out said method.
1 . A hardware accelerator, comprising:
an execution unit configured to execute at least one artificial intelligence, AI, model using one or more resources of the hardware accelerator;
a watermark verification unit configured to verify whether a predefined watermark is present in the AI model and to output a verification result indicative of the presence or absence of said watermark;
a resource management unit configured to enable said resources if the verification result indicates that the watermark is present; and
wherein the watermark verification unit is configured to verify whether the watermark is present during startup of the hardware accelerator, periodically during operation of the hardware accelerator, or upon request of a host processor.
2 . The hardware accelerator of claim 1 , wherein the resource management unit is configured to disable the resources if the verification result indicates that the watermark is absent.
3 . The hardware accelerator of claim 1 , wherein the resource management unit is further configured to block the execution unit from executing the AI model if the verification result indicates that the watermark is absent.
4 . The hardware accelerator of claim 1 , wherein the watermark comprises predefined pairs of inputs and expected outputs of the AI model.
5 . The hardware accelerator of claim 4 , wherein the expected outputs form part of a single output class.
6 . The hardware accelerator of claim 1 , wherein the resource management unit is further configured to adjust the execution speed and/or power profile of the execution unit in dependence on the verification result.
7 . The hardware accelerator of claim 1 , wherein the resource management unit is further configured to raise an alert if the verification result indicates that the watermark is absent.
8 . The hardware accelerator of claim 1 , wherein the watermark verification unit is configured to verify whether the predefined watermark is present in the AI model using protected data.
9 . The hardware accelerator of claim 8 , wherein the protected data comprises data protected by a cryptographic signature.
10 . A method of operating a hardware accelerator, the method comprising:
verifying, by a watermark verification unit comprised in the hardware accelerator, whether a predefined watermark is present in an artificial intelligence, AI, model and to output a verification result indicative of the presence or absence of said watermark;
enabling, by a resource management unit comprised in the hardware accelerator, one or more resources of the hardware accelerator if the verification result indicates that the watermark is present;
executing, by an execution unit comprised in the hardware accelerator, the AI model using said resources; and
wherein the resource management unit adjusts the execution speed and/or power profile of the execution unit in dependence on the verification result.
11 . The method of claim 10 , wherein the resource management unit disables the resources if the verification result indicates that the watermark is absent.
12 . The method of claim 10 , wherein the resource management unit blocks the execution unit from executing the AI model if the verification result indicates that the watermark is absent.
13 . The method of claim 10 , wherein the watermark comprises predefined pairs of inputs and expected outputs of the AI model.
14 . The method of claim 13 , wherein the expected outputs form part of a single output class.
15 . The method of claim 10 , wherein the watermark verification unit verifies whether the watermark is present during startup of the hardware accelerator, periodically during operation of the hardware accelerator, or upon request of a host processor.
16 . The method of claim 10 , wherein the resource management unit raises an alert if the verification result indicates that the watermark is absent.
17 . A non-transitory computer-readable medium comprising the computer program of claim 16 .
18 . A computer program comprising executable instructions which, when executed by a hardware accelerator, carry out the method of claim 10 .
19 . A hardware accelerator comprising:
an execution unit configured to execute at least one artificial intelligence, AI, model using one or more resources of the hardware accelerator;
a watermark verification unit configured to verify whether a predefined watermark is present in the AI model and to output a verification result indicative of the presence or absence of said watermark;
a resource management unit configured to enable said resources if the verification result indicates that the watermark is present; and
wherein the resource management unit is further configured to adjust the execution speed and/or power profile of the execution unit in dependence on the verification result.
20 . The hardware accelerator method of claim 19 , wherein the watermark verification unit verifies whether the watermark is present during startup of the hardware accelerator, periodically during operation of the hardware accelerator, or upon request of a host processor.