Secure ML Model Partitioning for Lower TEE Compute Load
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing technologies face high computational pressure in Trusted Execution Environments (TEE) due to the need to adapt different reasoning frameworks for various customer requirements, leading to increased development efforts.
Innovation Solution
A method involving splitting a machine learning model into a security application module, which performs encryption calculations in an isolated operating environment, optimizing and encrypting the data processing module using a heterogeneous accelerated compiler, and interacting with a custom operator in a Rich Execution Environment (REE) to reduce computational load on the TEE.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If different reasoning frameworks are adapted in the TEE to meet different customer requirements, then the system's adaptability is improved, but the computational pressure and development efforts in the TEE increase significantly
Solution Approach 1:
The patent segments the machine learning model into two parts: a security application module that runs in the TEE and a preset operator that runs in the REE. This segmentation allows the TEE to handle only the security-critical encryption calculations while the REE handles the framework-specific reasoning operations, thereby reducing the computational burden and development complexity in the TEE while maintaining adaptability to different customer requirements through the REE-side operators.
Solution Approach 2:
The patent introduces a preset operator as an intermediary component that bridges the TEE and REE. This operator receives input from the TEE's security application module, performs framework-specific reasoning operations in the REE, and returns results to the TEE. This intermediary approach enables the system to support multiple reasoning frameworks without requiring the TEE to adapt to each framework, thus reducing computational pressure and development efforts in the secure environment.
2Reliability
If all reasoning frameworks are integrated into the TEE to ensure security, then the security reliability is improved, but the computational pressure and energy consumption in the TEE increase
Solution Approach 1:
The patent segments the computational workload by separating security-critical operations from framework-specific operations. The security application module in the TEE handles only encryption and decryption operations, which are essential for security reliability. The preset operator in the REE handles the energy-intensive reasoning framework operations. This segmentation maintains security reliability while significantly reducing energy consumption in the TEE.
Solution Approach 2:
The preset operator serves as an intermediary that performs energy-intensive reasoning operations outside the TEE. By moving these operations to the REE, the system maintains security reliability through the TEE's encryption/decryption functions while reducing the energy consumption burden on the TEE hardware and battery-powered devices.
3Reliability
If the TEE handles all encryption calculations to ensure data security, then the security protection is improved, but the processing speed and efficiency decrease
Solution Approach 1:
The patent segments the processing pipeline to allow parallel execution of security operations and reasoning operations. The security application module in the TEE performs encryption/decryption while the preset operator in the REE performs reasoning calculations simultaneously. This segmentation improves processing speed by eliminating sequential dependencies while maintaining data security through the TEE's protected environment.
Data Source
AI summary
A data processing method, apparatus and system are disclosed. The method includes: splitting a preset data processing module from a machine learning model, to generate a security application module configured for performing encryption calculation on data input into the machine learning model; taking, through a preset operator, an output value of an operation layer in the split machine learning model as an input value to be input into the security application module, and inputting the input value into the security application module; performing, through the security application module, subgraph calculation in an isolated operating environment according to the input value, to obtain a calculation result; returning the calculation result to the preset operator. The present application solves the technical problem of high computational pressure of TEE model due to that different reasoning frameworks need to be adapted to different customer requirements in the TEE model in related technologies.