AI Model Provenance Verification with DLT Watermarks and Checksums
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods fail to effectively verify the ownership and integrity of AI models, particularly in public domains, and are vulnerable to tampering and unauthorized modifications.
Innovation Solution
A method using distributed ledger technology (DLT) to embed digital watermarks and checksums during AI model training, registering the model on a distributed ledger with unique IDs, and verifying ownership and integrity through checksum matching and watermark processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If digital watermarking is used to verify AI model ownership, then ownership verification capability is improved, but the method cannot detect model tampering or integrity violations
Solution Approach 1:
The patent combines digital watermarking with checksum verification into a unified integrity verification system. The watermark data is embedded in the AI model during training, and a checksum is generated from the watermark data. This merged approach allows simultaneous verification of both ownership (through watermark recognition) and integrity (through checksum validation), resolving the limitation where watermarking alone could not detect tampering.
Solution Approach 2:
The patent performs preliminary embedding of watermark data and generation of checksum values during the model training phase, before the model is deployed. This preliminary action ensures that the verification mechanisms are already in place and can immediately detect any subsequent tampering or unauthorized modifications, improving both ownership verification and integrity monitoring.
2Adaptability or versatility
If AI models are deployed in public domains, then model accessibility and utility are improved, but vulnerability to theft and unauthorized modification increases
Solution Approach 1:
The patent applies preliminary anti-action by embedding watermark data and establishing checksum verification mechanisms before the AI model is deployed in public domains. This preventive measure creates a built-in defense system that can detect and deter unauthorized modifications, allowing the model to be freely accessible while protecting against theft and tampering.
Solution Approach 2:
The patent introduces watermark data as an intermediary element that serves as a hidden identifier and integrity marker within the AI model. This intermediary allows verification of ownership and integrity without affecting the model's functionality or accessibility, enabling public deployment while maintaining security.
3Measurement precision
If comprehensive verification of AI model integrity is implemented, then detection accuracy of tampering is improved, but verification complexity and computational overhead increase
Solution Approach 1:
The patent extracts only the essential verification elements (watermark data and checksum values) from the complete AI model and stores them separately in a database. During verification, only these extracted elements need to be processed and compared, rather than analyzing the entire model. This extraction approach maintains high detection accuracy while significantly reducing verification complexity and computational overhead.
Data Source
AI summary
A method is for checking an integrity of an artificial intelligence (AI) model using distributed ledger technology (DLT). The method leverages state-of-the-art watermarking mechanism and ties it up with DLT to generate proof of origin (provenance) in a tamper-proof way. The AI model is registered on the distributed ledger (DL) by uploading a full checksum, a selective checksum, watermark data, and at least a predefined output of the watermark data. A unique model ID is received upon registration. The ownership and integrity of AI model is then determined by matching the model and the output of the watermark data followed by verification of the full checksum and the selective checksum of the AI model.