BARC talk by Nikita Kalinin

Friday, August 29, 2025, Nikita Kalinin, PhD student at the Institute of Science and Technology Austria, will give a talk on "Banded Inverse Matrices for Multi-Epoch Private Training with Correlated Noise".

Abstract:
DP-SGD is the standard way of differentially private training, where privacy is ensured by adding independent Gaussian noise at each iteration. Recent works show that correlating the noise across steps can significantly improve utility. Matrix factorization provides a principled way to correlate the noise. In the multi-epoch setting, most existing factorizations focus on making the correlation matrix banded. In this talk, I will argue that banding the inverse correlation matrix is a more natural and efficient choice, one that reduces factorization error, improves training accuracy, and is provably asymptotically optimal.

Bio:
Nikita Kalinin is a 4th-year PhD candidate and ELLIS PhD student at the Institute of Science and Technology Austria (ISTA), supervised by Christoph Lampert. His research focuses on differentially private model training, with a research interest in correlated noise methods.

Host:
Rasmus Pagh