Overview
Explore instance-hiding schemes for private distributed learning in this comprehensive seminar on theoretical machine learning. Delve into cryptographic techniques, data augmentation methods, and privacy-preserving algorithms as presented by Sanjeev Arora, a Distinguishing Visiting Professor from Princeton University. Examine topics such as Instahide, mixup data augmentation, multiplicative noise, and sine flip, while gaining insights into parameter mixing and experimental results. Analyze the implications of privacy laws, statistical indistinguishability, and practical attacks on privacy-preserving methods. Understand why mixup alone is not secure and discover the latest advancements in balancing machine learning efficiency with data privacy in distributed settings.
Syllabus
Introduction
Private Distributed Learning
Cryptography
Instahide
Mixup Data Augmentation
Cryptographic Instance Hiding
Multiplicative Noise
Sine Flip
Parameter Mixing
Experimental Results
Privacy Laws
Statistical Indistinguishability
Practical Attacks
Why Mixup Alone is Not Secure
Conclusion
Taught by
Institute for Advanced Study