Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
Learn about a novel approach to privacy-preserving machine learning through this Google TechTalk that introduces Differentially Private Prototype Learning (DPPL) as a solution for imbalanced transfer learning scenarios. Explore how traditional differential privacy methods like DP-SGD struggle in high privacy regimes with limited data and class imbalances, and discover how DPPL overcomes these limitations by leveraging publicly pre-trained encoders to extract features from private data. Understand the methodology of generating differentially private prototypes that represent each private class in embedding space while maintaining strong privacy guarantees under pure differential privacy. Examine how DPPL achieves high-utility predictions from minimal private training data without iterative noise addition, and learn about enhanced privacy-utility trade-offs through strategic use of public data beyond encoder pre-training. Review experimental results across four state-of-the-art encoders and four vision datasets that demonstrate DPPL's superior performance in challenging private learning setups with various data availability and class imbalance conditions.
