Third-party cyber evaluations involving OpenAI models
openai.com·13h ago
AI Research
arXiv CS.AI·May 26·5 min read
Extracting Training Data from Diffusion Language Models via Infilling
TL;DR
Existing methods for extracting training data from large language models primarily focus on prefix-conditioned extraction, which underestimates risks in diffusion language models (DLMs). A new method called infilling extraction was developed, allowing for data extraction using arbitrary binary masks.
✦ Why It Matters
Engineers and researchers should consider the implications of data extraction risks in diffusion language models for privacy and security.
Key Takeaways
- Implement infilling extraction to assess data memorization in DLMs.
- Use edge-conditioned masks for improved data extraction performance.
- Monitor for potential leakage of personally identifiable information.
- Adjust extraction parameters to optimize performance in DLMs.
How It Works
Infilling extraction utilizes a binary mask to selectively probe the memory of diffusion language models. Unlike prefix-conditioned methods, which only examine sequences from the start, this approach allows for denoising at any token position, leveraging the bidirectional nature of DLMs.
This flexibility enables the extraction of more comprehensive data, revealing hidden memorization patterns.