Neural Network Parameter Hash Authentication Across Framework Changes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural networks are vulnerable to malicious attacks that alter or replace them, compromising security-sensitive services such as authentication and payment, as integrity validation is lacking.
Innovation Solution
An electronic device generates a hash value based on the parameters of a neural network, comparing it with a pre-stored reference hash value to ensure integrity, allowing valid hash authentication even if the deep learning framework changes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If hash authentication is performed on the entire binary file of the neural network, then integrity validation is achieved, but the authentication fails when the deep learning framework type or version changes
Solution Approach 1:
The patent extracts the essential parameters of the neural network from the binary file, separating the framework-specific binary data from the framework-agnostic parameter data. By generating hash values based only on these extracted parameters rather than the entire binary file, the system achieves integrity validation that is independent of the deep learning framework type or version, thus resolving the contradiction between reliability and adaptability
Solution Approach 2:
The patent segments the neural network binary file into distinct components: framework-specific elements and framework-agnostic parameter elements. By applying hash authentication only to the parameter segment rather than the entire file, the system maintains reliability for the essential neural network functionality while allowing flexibility in the framework implementation details
2Productivity
If the neural network binary is updated frequently, then the application can use updated neural networks, but the reference hash value stored in the secure world becomes outdated
Solution Approach 1:
The patent extracts only the essential parameter information from the neural network binary, excluding framework-specific and version-specific binary data. This extraction allows the reference hash value to remain stable in the secure world while the application can update the neural network binary frequently, as long as the essential parameters remain unchanged, thus resolving the contradiction between productivity and reliability
3Object-affected harmful factors
If the neural network is altered or replaced by malicious attacks, then the application may offer incorrect services, but detecting such alterations requires complex integrity validation
Solution Approach 1:
The patent extracts the essential parameters that define the neural network's functionality from the binary file. By generating hash values from these extracted parameters rather than analyzing the entire complex binary structure, the system achieves effective malicious attack detection with reduced validation complexity. The extracted parameters serve as a simplified representation that maintains security effectiveness while reducing computational burden
Data Source
AI summary
An electronic device for performing hash authentication on a neural network, and an operating method therefor are disclosed. The disclosed electronic device comprises: a processor and a memory having at least one instruction executable by the processor. If at least one instruction is executed by the processor, the processor can be configured to: generate a hash value on the basis of parameters of a neural network used for an application to be executed, determine whether the generated hash value matches the pre-stored reference hash value in the security area of the processor, and approve the loading of the neural network in response to whether the generated hash value matches the reference hash value in the general area of the processor.