Virtual Machine Segmentation for Secure In-Vehicle ML
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current technologies for securing machine learning models in in-vehicle terminals face challenges in balancing security with increased calculation overheads and processing delays, such as encryption methods and trusted execution environments, which affect the performance of driver assistance functions.
Innovation Solution
The machine learning model is split into multiple virtual machines, each processing data from different sensor groups independently, with a fusion neural network combining outputs to provide driving parameter information, enhancing security by increasing the difficulty of model extraction and reducing calculation delays through mature interaction technologies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the machine learning model is encrypted to ensure security, then the security of the model is improved, but the calculation overheads increase and processing delay occurs
Solution Approach 1:
The machine learning model is segmented into multiple virtual machines, with each virtual machine containing a portion of the model. This segmentation allows the model to run distributedly without requiring encryption of the entire model, thereby maintaining security through distribution while avoiding the calculation overhead associated with encrypting and decrypting the full model during inference.
2Reliability
If the machine learning model is stored in a trusted execution environment (TEE), then the security of the model is improved, but the computing power and space limitations of TEE cause increased interaction frequency and processing delay
Solution Approach 1:
The model is segmented across multiple virtual machines rather than confined to a single TEE environment. This approach distributes the computational load and reduces the frequency of interactions with the TEE, thereby improving calculation efficiency while maintaining security through the distributed architecture.
Solution Approach 2:
The system allows the machine learning model to run in a general environment across multiple virtual machines rather than being restricted to the specialized TEE environment. This multi-functionality approach enables the model to leverage broader computational resources while security is maintained through the distributed virtual machine architecture.
3Reliability
If the machine learning model is encrypted and decryption is performed in each calculation step, then the security of the model is improved, but the calculation overheads increase significantly
Solution Approach 1:
The model is divided into segments distributed across virtual machines, eliminating the need for repeated encryption and decryption operations during calculation. Each virtual machine processes its portion of the model in plaintext, significantly reducing calculation overheads while security is maintained through the distributed architecture that prevents any single point of compromise.
Data Source
AI summary
A data processing apparatus for an intelligent vehicle has a plurality of virtual machines disposed therein, and a first neural network model corresponding to a sensor group in a machine learning model is disposed in one virtual machine. The machine learning model performs calculation on detection data of a corresponding sensor group in an independent virtual machine. Subsequently, all first neural network models in the plurality of virtual machines send output results of detection data of a plurality of sensor groups to a second neural network model, and the second neural network model obtains, based on the plurality of output results, a fusion output result used to indicate vehicle driving parameter information.