701: Generative A.I. without the Privacy Risks — with Prof. Raluca Ada Popa

Privacy Safeguarding

Ensuring privacy while using AI APIs like GPT-4 is crucial, and Raluca Ada Popa offers innovative solutions. She explains how Opaque Systems is developing a method to strip personally identifiable information (PII) from queries before sending them to AI models, replacing sensitive data with symbolic information 1. This approach allows AI to process queries without accessing private data, ensuring compliance with regulations like HIPAA 2.

From the user's perspective and experience is the same as if they were talking to GPT four. But the difference is that GPT four doesn't get to see any PIi data, they only see symbols.

--- Raluca Ada Popa

This method maintains the functionality of AI while safeguarding user privacy.

   

Open Source Privacy

Open-source models offer distinct privacy advantages over proprietary ones, according to Raluca Ada Popa. She highlights that open-source models can be run on-premises, allowing organizations to keep sensitive data in-house, unlike proprietary models that require sending data to external providers 3. However, challenges remain in managing access control within organizations to prevent unauthorized data access 3.

There's a clear advantage of open source models over proprietary ones when it comes to privacy and confidentiality.

--- Raluca Ada Popa

Balancing open and closed-source models involves trade-offs between innovation and security.

   

Data Protection Challenges

Protecting data privacy in AI, especially with LLMs, presents significant challenges. Raluca Ada Popa discusses the need for domain-specific LLMs that are fine-tuned for particular industries, enhancing their effectiveness while maintaining confidentiality 4. She also emphasizes the importance of setting advanced data policies to control data exposure, ensuring that only necessary information is shared 5.

You can set very, very advanced policies.

--- Raluca Ada Popa

These strategies are crucial for balancing data utility and privacy.