LLM Marking via Training Data Triplet Embedding

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Large language models (LLMs) are expensive to develop and are at risk of being appropriated by third parties, with existing watermarking techniques being easily removable and falsifiable.

Innovation Solution

A method for marking LLMs using an identification database of triplets comprising prompts, input keys, and output keys, integrated during training, allowing for subsequent model identification through user interaction.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If watermarking techniques are used to mark the model, then model identification is enabled, but the marking can be easily removed and falsified

Engineering Contradiction:
Improvemodel identification reliabilityVSAvoidvulnerability to removal and falsification
Core Design Contradiction:
ReliabilityVSObject-affected harmful factors

Solution Approach 1:

The patent embeds identification data during the model training phase rather than adding watermarks after training. This preliminary action integrates the marking process into the fundamental structure of the model, making it inseparable from the model's core functionality and resistant to removal or falsification attempts.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent combines the model training process with the marking process by integrating identification data into the training dataset. This merging ensures that the identification capabilities are built into the model's weight parameters and cannot be easily separated or tampered with independently.

Inventive Principle:
Principle #5Merging (Combining)

2Reliability

If the model is marked during training, then the marking is difficult to falsify, but additional training time and resources are required

Engineering Contradiction:
Improvemarking authenticityVSAvoidtraining time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent makes the training dataset serve dual purposes: it trains the model for its primary function while simultaneously embedding identification data for marking. This multi-functionality eliminates the need for separate marking operations and reduces overall training time despite the added reliability of integrated marking.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Solution Approach 2:

The patent incorporates identification data as a portion of the overall training dataset rather than requiring complete retraining or additional extensive training. This partial action approach achieves reliable marking while minimizing the time and resource overhead associated with the marking process.

Inventive Principle:
Principle #16Partial or excessive action

3Reliability

If identification data is embedded in the model, then ownership proof is secured, but the model architecture and performance may be affected

Engineering Contradiction:
Improveownership identificationVSAvoidmodel architecture complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent embeds identification data in specific local regions of the training dataset rather than uniformly modifying the entire model architecture. This localized approach ensures that identification capabilities are integrated without significantly altering the overall model structure or performance characteristics.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent modifies the training dataset parameters to include identification data, which then influences the model's weight parameters during training. This parameter-based approach allows for ownership identification without requiring fundamental changes to the model's architectural design.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4557130A1Method for marking and identifying large language model
Publication Date: 2025.05.21 THALES SA
  • EP4557130A1 patent drawingFigure 1
  • EP4557130A1 patent drawingFigure 2
  • EP4557130A1 patent drawingFigure 3

AI summary

The present invention relates to a method for marking a large language model, called an LLM model, the LLM model being capable of generating output data following the obtaining as input to the model: of a textual command, called a prompt, and of an input data, the marking method comprising: - obtaining an identification database, each identification data being a triplet comprising a prompt, an input key and an output key, and - marking the LLM model with the identification data so that, for each identification data, the LLM model generates the output key of the identification data following the obtaining as input to the LLM model: of the prompt of the identification data and of the input key of the identification data as input data.