Character Encoding and Signed Manifests for Generative-Model Provenance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods struggle to accurately track the provenance of digital content generated by generative models, particularly in documents edited by users, as watermarking and manifest techniques can be lost or misinterpreted during editing, and classifiers become less effective with improved generative models.
Innovation Solution
A computing system with a provenance determination module that uses character encoding and digitally signed manifests to identify model-generated content, providing visual indications and tracking edit operations, including generative model attribution.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If watermarking and manifest techniques are used to track provenance, then model-generated content can be identified, but the provenance metadata can be lost or misinterpreted during editing operations
Solution Approach 1:
The provenance tracking is segmented into two independent components: (1) character-level encoding that marks individual generated characters with provenance information, and (2) a manifest file that stores metadata about the generation process. This segmentation allows the character encoding to survive editing operations while the manifest provides comprehensive provenance information.
Solution Approach 2:
Provenance information is embedded into the character encoding at the time of content generation, before any editing operations occur. This preliminary embedding ensures that the provenance markers are already in place and can withstand subsequent editing operations without being lost or misinterpreted.
2Measurement precision
If classifiers are used to detect model-generated content, then provenance can be determined, but effectiveness decreases as generative models improve
Solution Approach 1:
Instead of relying on classifiers to detect model-generated content after generation, the system embeds provenance information directly into the character encoding during the generation process itself. This preliminary action ensures that detection remains accurate regardless of improvements in generative model capabilities, as the provenance markers are built-in rather than detected through pattern recognition.
3Ease of operation
If visual indications are added to show model-generated content, then user awareness is improved, but device complexity increases
Solution Approach 1:
The system introduces a manifest file as an intermediary that stores comprehensive provenance information separately from the edited content. This intermediary allows the visual indication layer to remain simple (just displaying markers in the text) while the complex provenance tracking logic is encapsulated in the manifest processing, effectively separating the user-facing simplicity from the backend complexity.
Data Source
AI summary
A computing system is provided that includes processing circuitry and associated memory. The processing circuitry is configured to implement a program using portions of the associated memory, to receive, via an edit operation, digital content and provenance metadata associated with the digital content. The processing circuitry implementing the program is further configured to determine, via a provenance determination module, that a textual portion of the digital content is model-generated and originated from a generative model, based on the provenance metadata, and output the digital content to a graphical user interface with a visual indication that the textual portion of the digital content is model-generated.