Google, IBM & Meta Certificates – 40% Off
One plan covers every Professional Certificate on Coursera.
Unlock All Certificates
Learn about the theoretical foundations and practical applications of one-run auditing for differential privacy in this Google TechTalk. Explore how this auditing approach can efficiently quantify privacy leakage by identifying multiple elements in a single mechanism run, rather than requiring many separate runs for single-element identification. Discover the conditions under which one-run auditing achieves asymptotic tightness and understand how interference between observable effects of different data elements affects privacy estimation accuracy. Examine the relationship between approximately local mechanisms and the effectiveness of one-run auditing, including how many common mechanisms can be adapted to work within this framework. Investigate adaptive variants that reduce interference through sequential element participation guessing and learn about enhanced approaches for DP-SGD that leverage full learned weights via KKT-based certificates and influence-function scores to enable more informative multi-dimensional participation tests.
