Security optimizing compute distribution in a hybrid deep learning environment

Patent metadata

Field	Value
Publication number	US-11574175-B2
Application number	US-202016912152-A
Country	US
Kind code	B2
Filing date	Jun 25, 2020
Priority date	Jun 25, 2020
Publication date	Feb 7, 2023
Grant date	Feb 7, 2023

Abstract

Embodiments are directed to security optimizing compute distribution in a hybrid deep learning environment. An embodiment of an apparatus includes one or more processors to determine security capabilities and compute capabilities of a client machine requesting to use a machine learning (ML) model hosted by the apparatus; determine, based on the security capabilities and based on exposure criteria of the ML model, that one or more layers of the ML model can be offloaded to the client machine for processing; define, based on the compute capabilities of the client machine, a split level of the one or more layers of the ML model for partition of the ML model, the partition comprising offload layers of the one or more layers of the ML model to be processed at the client machine; and cause the offload layers of the ML model to be downloaded to the client machine.

First claim

What is claimed is:

1. An apparatus comprising:
one or more processors to:
determine security capabilities and compute capabilities of a client machine requesting to use a machine learning (ML) model hosted by the apparatus;
determine, based on the security capabilities and based on exposure criteria of the ML model, that one or more layers of the ML model can be offloaded to the client machine for processing, wherein the exposure criteria comprises identification of a model protection level (MPL) that is a lowest layer of the one or more layers of the ML model that can be offloaded to the client machine without exposing the ML model to a security risk, the lowest layer counted from a first input layer of the one or more layers of the ML model;
define, based on the compute capabilities of the client machine, a split level of the one or more layers of the ML model for partition of the ML model, the partition comprising offload layers of the one or more layers of the ML model to be processed at the client machine; and
cause the offload layers of the ML model to be downloaded to the client machine.

2. The apparatus of claim 1 , wherein the one or more processors to determine that the one or more layers of the ML model can be offloaded further comprises the one or more processors to compare the exposure criteria to the security capabilities, wherein the security capabilities identify a data protection level (DPL) of the one or more layers that can run on the client machine to ensure that confidential data of the client machine is not exposed.

3. The apparatus of claim 2 , wherein offload of the one or more layers to the client machine is allowed in response to the DPL being less than or equal to the MPL.

4. The apparatus of claim 1 , wherein the one or more processors to define the split level further comprises the one or more processors to identify the split level based on a level of compute power used to run a subset of the one or more layers defined by the split level being less than or equal to a compute power defined in the compute capabilities of the client machine.

5. The apparatus of claim 1 , wherein the one or more layers of the ML model further comprise a set of client machine layers that are trained by the client machine and are run by the client machine, where client machine layers at least one of replace the offload layers of the ML model or are run in addition to the offload layers of the ML model.

6. The apparatus of claim 5 , wherein training of the client machine layers of the ML model comprises back-propagating vectors to adjust first weights in the client machine layers without adjusting second weights in a common computing core of the ML model.

7. The apparatus of claim 6 , wherein an inference stage of the ML model utilizes the client machine layers and the common computing core of the ML model, and wherein an inference signal is generated by processed by starting processing at the client machine layers, proceeding to processing at the common computing core, and returning to processing at the client machine layers to provide an inference result.

8. The apparatus of claim 1 , wherein the one or more layers of the ML model comprise one or more stages of a multifunction perceptron architecture having a plurality of neurons to perform one or more neuron functions of the ML model, and wherein the plurality of neurons comprise heterogenous neurons including one or more of splitter neurons, mixer neurons, counter neurons, selector neurons, extractor neurons, or transformer neurons.

9. The apparatus of claim 8 , wherein the split level for the partition of the ML model is defined in terms of the one or more stages of the multifunction perceptron architecture and identifies at least one stage of the ML model for offload to the client machine for at least one of training or inference of the ML model.

10. The apparatus of claim 1 , wherein the one or more processors comprise one or more of a graphics processor, an application processor, and another processor, wherein the one or more processors are co-located on a common semiconductor package.

11. A non-transitory computer-readable storage medium having stored thereon executable computer program instructions that, when executed by one or more processors, cause the one or more processors to perform operations comprising:
determining security capabilities and compute capabilities of a client machine requesting to use a machine learning (ML) model;
determining, based on the security capabilities and based on exposure criteria of the ML model, that one or more layers of the ML model can be offloaded to the client machine for processing, wherein the exposure criteria comprises identification of a model protection level (MPL) that is a lowest layer of the one or more layers of the ML model that can be offloaded to the client machine without exposing the ML model to a security risk, the lowest layer counted from a first input layer of the one or more layers of the ML model;
defining, based on the compute capabilities of the client machine, a split level of the one or more layers of the ML model for partition of the ML model, the partition comprising offload layers of the one or more layers of the ML model to be processed at the client machine; and
causing the offload layers of the ML model to be downloaded to the client machine.

12. The non-transitory computer-readable storage medium of claim 11 , wherein the operations further comprising determining that the one or more layers of the ML model can be offloaded further comprises the one or more processors to compare the exposure criteria to the security capabilities, wherein the security capabilities identify a data protection level (DPL) of the one or more layers that can run on the client machine to ensure that confidential data of the client machine is not exposed.

13. The non-transitory computer-readable storage medium of claim 11 , wherein the one or more layers of the ML model further comprise a set of client machine layers that are trained by the client machine and are run by the client machine, where client machine layers at least one of replace the offload layers of the ML model or are run in addition to the offload layers of the ML model, and training of the client machine layers of the ML model comprises back-propagating vectors to adjust first weights in the client machine layers without adjusting second weights in a common computing core of the ML model.

14. The non-transitory computer-readable storage medium of claim 13 , wherein an inference stage of the ML model utilizes the client machine layers and the common computing core of the ML model, and wherein an inference signal is generated by processed by starting processing at the client machine layers, proceeding to processing at the common computing core, and returning to processing at the client machine layers to provide an inference result.

15. The non-transitory computer-readable storage medium of claim 11 , wherein the one or more layers of the ML model comprise one or more stages of a multifunction perceptron architecture having a plurality of neurons to perform one or more neuron functions of the ML model, and wherein the split level for the partition of the ML model is defined in terms of the one or more stages of the multifunction perceptron architecture and identifies at least one stage of the ML model for offload to the client machine for at least one of training or inference of the ML model.

16. A method comprising:
determining security capabilities and compute capabilities of a client machine requesting to use a machine learning (

Assignees

• Intel Corp

Inventors

• Pogorelik Oleg
• Nayshtut Alex
• Kounavis Michael E
• Kellermann Raizy
• Durham David M

Classifications

• G06N3/09

  Supervised learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/098

  Distributed learning, e.g. federated learning · CPC title

• G06F21/6245 Primary

  Protecting personal data, e.g. for financial or medical purposes · CPC title

• G06N3/084

  Backpropagation, e.g. using gradient descent · CPC title