A Dive into Privacy-Preserving Machine Learning

In this technical exploration from the DEF CON Safe Mode AI Village, researchers Daniel and Farhad deliver "Secrets Are Lies, Sharing Is Caring, Privacy Is Theft," examining privacy vulnerabilities in machine learning architectures. Opening at 00:13, the speakers address how trained models frequently leak sensitive training data through output statistics and prediction probabilities.

At 01:00 and 02:00, the presentation examines the delicate balance between machine learning utility and data privacy, evaluating how much model accuracy must be traded off to achieve robust data confidentiality. At 03:00 and 04:00, the authors review classic linkage attacks, citing Latanya Sweeney's landmark re-identification of medical records and the de-anonymization of the Netflix Prize dataset.

The speakers explore modern privacy-preserving frameworks past 10:00, detailing the mathematical mechanics of Differential Privacy, homomorphic encryption, and secure multi-party computation to protect models against membership inference attacks. (AI summary — Gemini Flash)