Neural Network Watermarking for Authentication and Unauthorized Use Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural networks lack mechanisms to prevent unauthorized use and theft, and there is a need to distinguish between authorized and unauthorized generation of neural networks to protect users from potential damage.
Innovation Solution
A technique is introduced to embed information into neural networks, such as generation source information, authentication information, and non-fungible tokens, which can be extracted to authenticate authorized use, ensuring only authorized users can utilize the networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If information embedding mechanisms are added to neural networks, then security and authentication capability are improved, but device complexity increases
Solution Approach 1:
The patent embeds authentication information directly within the neural network's internal structure (weights, biases, or activation functions), nesting the security mechanism inside the existing network architecture rather than adding external authentication layers. This allows security functionality to be integrated without significantly increasing overall system complexity.
Solution Approach 2:
The patent combines multiple functions into unified components: authentication information is merged with weight parameters, generation source information is combined with network structure, and verification processes are integrated into the forward propagation routine. This merging reduces the need for separate security subsystems.
2Reliability
If authentication information is embedded in neural network elements, then unauthorized use prevention is improved, but manufacturing precision requirements increase
Solution Approach 1:
The patent embeds authentication information in specific localized regions of the neural network (certain weight matrices, bias vectors, or activation functions) rather than uniformly across the entire network. This localized embedding reduces the precision requirements for the entire network while maintaining security in critical areas.
Solution Approach 2:
The patent uses parameter transformation techniques where authentication information is encoded into weight parameters through reversible mathematical transformations. This allows precise embedding of security information while maintaining the statistical properties needed for network functionality, reducing overall precision requirements.
3Reliability
If embedded information is added to neural networks, then traceability and authentication are improved, but loss of information in the network increases
Solution Approach 1:
The patent extracts authentication and traceability information as separate metadata that can be independently verified without affecting the core neural network computations. This extracted information is stored in dedicated fields or separate data structures, preventing interference with the network's primary information processing functions.
Solution Approach 2:
The patent introduces intermediary verification mechanisms that check embedded information without requiring modification of the main computational flow. These intermediaries act as mediators between the embedded security information and the verification process, preventing information loss in the primary network operations.
Data Source
AI summary
Provided an information processing system including: an embedded information acquisition unit which acquires embedded information; and a generation unit which generates an embedded neural network in which the embedded information is embedded into an element of a neural network.