Private Generative Models: Balancing Privacy and Utility in AI
summary
Private generative models claim to protect individuals while keeping data useful by using techniques like differential privacy, teacher‑student aggregation, federated learning and generative families (GANs, VAEs, diffusion). Concrete failures include GANs leaked health data and API keys and the New York taxi reidentification via flawed anonymization. Experiments show adding noise cuts utility — an Iris dataset example visualizes the privacy/utility tradeoff — so teams measure with privacy metrics, sanity checks and statistical overlap tests to decide the right balance.
AI-written summary · about
chapters (9)
- 0:00Introduction and expectations
- 1:11Generative vs discriminative models
- 6:45How generative models create data
- 11:49Applications of generative AI
- 13:56Privacy risks in generative models
- 18:19Methods for preserving privacy
- 23:41Balancing privacy and utility
- 25:31Measuring privacy and utility
- 26:37Recap and conclusion
related talks
45mCombating Generative AI's Privacy Abuses
28mPrivacy preserving machine learning
13mNavigating Generative AI and Its Risks
14mBuilding Transparent and Ethical Generative AI Models in High-Stakes Industries
31mIs your model private?
01h 20mUnlocking societal benefits of AI with privacy protective technologies.
54mOpen Data vs. Privacy - How Synthetic Data Can Democratize AI Innovation
15mNavigating the Privacy Landscape of AI on the Devices: Challenges and Best Practices