The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Ensuring Privacy for Any LLM with Patricia Thaine - #716
71 snips
Jan 28, 2025 
Patricia Thaine, co-founder and CEO of Private AI, specializes in privacy-preserving AI techniques. She dives into the critical issues of data minimization, the risks of personal data leakage from large language models (LLMs), and the challenges of redacting sensitive information across different formats. Patricia highlights the limitations of data anonymization, the balance between real and synthetic data for model training, and the evolving landscape of AI regulations like GDPR. She also discusses the ethical considerations surrounding bias in AI and the future of privacy in technology.
AI Snips
Chapters
Transcript
Episode notes
Internal Model Risks
- Internally hosted models still pose data breach risks.
- Data minimization limits the impact of potential breaches.
Data Leakage
- Models can memorize and leak training data, including from embeddings.
- Embeddings, even if not directly invertible, can leak sensitive information like salaries.
Embedding Leakage
- Embeddings can leak data both through direct attacks and through model outputs.
- Sensitive information in embedding databases can be revealed in model responses.