Third-party cyber evaluations involving OpenAI models
openai.com·13h ago
AI Research
arXiv CS.AI·Jun 10·5 min read
Local Is Not a Sufficient Privacy Boundary: Governing OS-Integrated On-Device AI
TL;DR
As AI systems integrate into operating systems, privacy concerns extend beyond local computation to include data assembly and context management. A new privacy framework was developed that includes a threat model, a six-part risk taxonomy, and an audit rubric for on-device AI.
✦ Why It Matters
Engineers can implement this framework to improve privacy governance in their on-device AI systems.
Key Takeaways
- Adopt a comprehensive privacy framework for AI systems.
- Implement a six-part risk taxonomy to assess privacy risks.
- Ensure user control and visibility in AI applications.
- Conduct regular audits to maintain privacy standards.
Full Summary
With the rise of on-device AI, traditional notions of privacy are challenged, as local processing does not guarantee data protection. The authors propose a comprehensive privacy framework that addresses these challenges by treating privacy as an institutional accountability issue rather than merely a technical feature.
This framework includes a detailed threat model, a six-part privacy risk taxonomy, and privacy-by-architecture controls, along with a four-level audit rubric. They applied this rubric to compare Apple Intelligence/Foundation Models, Android AICore/Gemini Nano, and Microsoft Recall, revealing significant differences in privacy practices.
Key findings indicate that effective privacy requires constrained information flow, limited authority, visible user control, and auditable governance. These insights are crucial for developers and researchers aiming to enhance privacy in AI applications.