Learning from private data Algorithms for decentralized and differentially private machine learning

The growth of artificial intelligence creates a demand for data, yet the majority of the world's most valuable information remains behind organizational firewalls and in private silos. Public datasets enable advances in machine learning, but proprietary data from enterprises, institutions, governments, and individuals represents an untapped data reservoir for the next generation of AI – if only this data could be accessed without compromising privacy.
This thesis presents advancements to the theory and practice of private learning systems.
Rather than treating privacy as a constraint that diminishes model quality, we establish how the designed algorithms shift the risk-reward balance and make data sharing feasible where it was previously considered unfeasible. Our work spans several domains in which privacy concerns hinder effective data utilization.
In the context of pandemic response, we show that contact tracing systems can achieve pandemic control while having both communication efficiency and privacy guarantees. We again apply this privacy lens to develop new methods in the field of deep learning, modern large language models, and private in-context learning for small local datasets.
Our contributions demonstrate the interplay between utility and privacy-preserving learning. Many of our works improve the trade-off or present advantages over existing methods. Ultimately, we show a path toward machine learning that integrates previously unused data. In addition to theoretical aspects and computational efficiency, we provide both the foundations and compelling examples to expand the frontier of machine learning beyond public datasets.