Third-party cyber evaluations involving OpenAI models
openai.com·13h ago
AI Research
arXiv CS.AI·Jun 10·5 min read
IDP-Bench: Benchmarking ability of LLMs to protect personal information in interdependent privacy contexts
TL;DR
Privacy concerns arise as large language models (LLMs) handle sensitive user data, particularly in interdependent privacy (IDP) contexts where one person's data can be exposed through others. IDP-Bench was developed to benchmark LLMs' understanding of IDP scenarios using the Contextual Integrity framework, evaluating eight open-source models.
✦ Why It Matters
Engineers can leverage IDP-Bench to enhance LLM privacy features and address interdependent privacy challenges.
Key Takeaways
- Utilize IDP-Bench to evaluate LLM privacy capabilities.
- Focus on improving identification of CI parameters in models.
- Consider model size when assessing IDP understanding.
- Conduct targeted research on interdependent privacy in AI.
Full Summary
As large language models (LLMs) are increasingly used as personal assistants, privacy becomes a critical issue, especially in interdependent privacy (IDP) contexts where individuals' data can be unintentionally revealed by others. To address this, IDP-Bench was created as the first benchmark specifically for evaluating LLMs' capabilities in IDP scenarios, grounded in the Contextual Integrity (CI) framework.
Eight open-source LLMs were assessed across three levels of IDP reasoning using two LLM judges. Findings revealed that while six out of eight models recognized co-ownership of data with over 90% accuracy, they struggled with identifying essential CI parameters, such as information attributes and secondary subjects, scoring below 74% in most cases.
Additionally, five out of eight models scored below 77% in judging the appropriateness of data sharing. The study indicates that larger models perform better in understanding IDP, but smaller models often underperform, emphasizing the need for targeted research in this area.