Auditing unauthorized training data from AI generated content using information isotopes

Area of Science:

• Artificial Intelligence
• Data Science
• Computer Science

Background:

• Rapid AI growth relies on vast human data, raising concerns about intellectual property and privacy violations.
• Auditing AI misuse is difficult due to black-box systems that hide training and inference processes.

Purpose of the Study:

• To introduce a novel method for tracing training data within opaque AI systems.
• To provide a framework for detecting and evidencing data utilization in AI models.

Main Methods:

• Inspired by chemical isotope tracing, the study proposes 'information isotopes' to mark and track data.
• A framework selectively marks target data elements and detects their propagation in AI model outputs.
• The method enables black-box auditing of data usage.

Main Results:

• Experiments on thirteen AI models across six datasets show high accuracy (up to 99%) in distinguishing training from non-training data.
• Results demonstrate strong statistical significance (p < 0.01) with approximately 4,000 words of evidence.
• The approach provides concrete evidence of data utilization.

Conclusions:

• The information-isotope tracing framework offers a practical solution for protecting data rights in AI.
• An open-source tool is available to support the implementation of this data tracing method.
• This work addresses critical challenges in AI transparency and data governance.