Neural Network Verification via Hash Code Comparison
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
There is a need for a simple and cost-effective method to secure and verify neural networks to prevent or detect unauthorized access and modifications, as existing methods are either complex or inadequate in ensuring data integrity and authenticity.
Innovation Solution
A method involving the calculation of hash codes using a collision-resistant hash function for neural networks, where the hash code is stored in a secure database with an identifier, allowing for verification by comparing the calculated hash code of a neural network with the stored one, thereby ensuring the network's integrity and authenticity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a neural network is made available to other users for use, then the neural network can be utilized and its performance characteristics can be applied, but the risk of unauthorized access and modifications increases
Solution Approach 1:
The patent applies preliminary action by calculating and storing a hash code of the neural network's performance characteristics before the network is made available to users. This pre-computed hash code serves as a reference value that enables future verification of the network's integrity without requiring additional protective measures during distribution. The hash code is stored in association with the neural network in a database, creating a baseline for detecting any subsequent unauthorized modifications.
Solution Approach 2:
The patent introduces a hash code as an intermediary element that mediates between the neural network and the verification process. Instead of directly protecting the complex neural network structure, the system uses this simplified hash representation as a mediator to verify integrity. The hash code acts as a fingerprint that can be easily computed and compared, serving as an intermediate verification mechanism that protects the underlying neural network from unauthorized modifications while allowing legitimate use.
2Reliability
If complex security measures are implemented to protect neural networks from unauthorized modifications, then security is improved, but the complexity and cost of the system increases
Solution Approach 1:
The patent extracts the essential verification information from the complex neural network by computing a hash code based on the network's performance characteristics. Instead of attempting to protect or verify the entire complex neural network structure, the system extracts a simplified numerical representation (hash code) that captures the essential integrity information. This extracted hash code can be easily stored, transmitted, and compared, significantly reducing the complexity of the security system while maintaining reliable verification capability.
Solution Approach 2:
The patent transforms the complex neural network parameters (weights, biases, architecture) into a simplified hash code parameter. This parameter change converts the verification problem from comparing complex neural network structures to comparing simple numerical hash values. The hash function acts as a transformation mechanism that maps the high-dimensional neural network parameter space into a low-dimensional hash space, making verification computationally efficient and system complexity manageable.
3Measurement precision
If the entire neural network data is stored and compared for verification, then verification accuracy is improved, but storage requirements and processing time increase
Solution Approach 1:
The patent creates a simplified copy of the neural network's essential characteristics in the form of a hash code. Instead of storing and comparing the entire neural network data (weights, biases, architecture), the system stores a compact hash representation that serves as a verification copy. This hash code copy contains sufficient information to verify integrity while occupying minimal storage space and enabling rapid comparison operations.
Solution Approach 2:
The patent segments the verification process into two distinct parts: (1) computing a hash code from the neural network's performance characteristics, and (2) comparing this hash code against a stored reference value. This segmentation separates the complex neural network data from the verification operation, allowing the system to verify integrity using only the compact hash code rather than requiring access to or storage of the complete neural network parameters.
Data Source
AI summary
The application relates to a method for verifying characteristic features of a neural network, comprising obtaining the neural network as well as an identifier assigned to the neural network, determining the characteristic features of the neural network, calculating a first hash code using a predetermined hash function from the characteristic features of the neural network, obtaining a second hash code assigned to the identifier from a secure database, as well as verifying the neural network by comparing the first hash code to the second hash code. The application furthermore comprises a computer software product which can be downloaded to the internal memory of a digital computer and which comprises software code sections with which the steps according to the method described here are carried out when the software is executed on a computer.