Method and apparatus for remote verification of artificial intelligence models

Remote Verification Technology for AI Models on Mobile Platforms Using TEE-Based Result Comparison

This technology is a method and apparatus for remotely verifying the integrity of an AI model on a mobile platform. The server transmits a target image to the mobile platform, receives the result generated by the platform's deep learning model, and compares it with the result from a second deep learning model that shares the same initial parameters.

AI models embedded in mobile platforms are vulnerable to parameter tampering attacks, which can lead to manipulated object recognition results. Previously, there was a lack of technical means to remotely verify and defend against such threats.

This technology proposes a method to ensure the security of the verification process itself by performing image transmission and result reception within a Trusted Execution Environment (TEE) isolated from the open operating system. It can be applied to security monitoring for autonomous vehicles and unmanned mobile platforms, enabling remote detection of parameter tampering and ensuring the integrity of AI models.

‍

‍

Key Features:

• The server transmits a target image, which is at least one image used for object recognition, to the mobile platform.
• The server receives the result from the mobile platform, which is obtained by applying the target image to the deep learning model.
• A second deep learning model is configured to have the same initial parameters as those applied to the deep learning model on the mobile platform.
• The mobile platform's deep learning model is verified by comparing the received result with the second result generated by the second deep learning model.

‍

This invention was developed with the support of the Ministry of Science and ICT's research project for the development of common core security technologies for unmanned mobile platforms.