Privacy in AI

Achieving reliable performance in modern AI systems deployed in sensitive domains such as healthcare requires genuine in-domain data, which is often absent from standard web-scale corpora. The use of such data imposes strict ethical and legal privacy requirements. The gold standard notion of differential privacy provides rigorous, semantically meaningful guarantees. However, enforcing such guarantees in foundation models raises new challenges of maintaining utility at the scale and computational cost of modern large models. We study the theory and practice of privacy-preserving machine learning in this large-scale regime. We design algorithms for training and inference in modern AI systems under differential privacy constraints. We analyse trade-offs between privacy, utility, and computational cost, and characterise fundamental statistical and computational limits. A central objective is to make large-scale generative models provably trustworthy when deployed in settings where data confidentiality is essential.