On-Device AI Model Parameter Protection via TEE Partitioning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
On-device artificial intelligence (AI) model parameters are vulnerable to copying and unauthorized use, and releasing them poses privacy risks due to exposure of training data information.
Innovation Solution
Implementing a trusted execution environment (TEE) on electronic devices to run AI models, where AI model parameters are protected by partitioning inference calculations between internal and external processor resources within and outside the TEE, ensuring secure execution and obfuscation of model data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If AI models are run locally on user devices, then user data privacy is improved and communication costs are reduced, but AI models are exposed to copying and unauthorized use
Solution Approach 1:
The patent segments the AI model execution into two distinct environments: a Trusted Execution Environment (TEE) for secure model parameter storage and processing, and a non-TEE environment for application logic. This segmentation isolates sensitive model data within the TEE while allowing controlled access through defined interfaces, preventing unauthorized copying while maintaining functionality.
Solution Approach 2:
The patent introduces a TEE as an intermediary layer between the AI model and the external environment. This intermediary provides a secure sandbox that protects model parameters from direct access while still enabling inference operations through controlled data exchange mechanisms, thus balancing privacy protection with model usability.
2Productivity
If AI model parameters are released for processing, then computation efficiency is improved, but privacy risks increase due to exposure of training data information
Solution Approach 1:
The patent segments the computation process into internal calculations performed within the TEE and external calculations performed outside the TEE. Sensitive operations involving model parameters remain inside the TEE, while less sensitive tasks can be offloaded externally. This segmentation enables efficient computation while maintaining privacy by ensuring that training data information never leaves the secure environment.
Solution Approach 2:
The patent applies different security qualities to different parts of the computation system. The TEE environment provides enhanced security protections for operations involving model parameters, while the external environment operates with standard security measures. This localized security approach allows efficient computation where needed without compromising privacy where sensitive data is handled.
3Reliability
If all calculations are performed within TEE, then model parameter security is improved, but processing performance deteriorates
Solution Approach 1:
The patent applies partial action by performing only the necessary calculations within the TEE boundary rather than all calculations. Critical operations involving model parameters remain inside the TEE for security, while peripheral or less sensitive calculations are performed outside the TEE to leverage additional processing resources and improve overall performance.
Data Source
AI summary
An electronic device includes at least one transceiver, at least one memory, and at least one processor coupled to the at least one transceiver and the at least one memory. The at least one processor is configured to receive, via the at least one transceiver, an AI model in a trusted execution environment (TEE). The at least one processor is also configured to receive an inference request and input data from a source outside the TEE. The at least one processor is further configured to partition a calculation of an inference result between an internal calculation performed by processor resources within the TEE and an external calculation performed by processor resources outside the TEE. In addition, the at least one processor is configured to produce the inference result based on results of the internal calculation and the external calculation.