M54

Privacypreserving logistic regression
Kamalika Chaudhuri and Claire Monteleoni

Problem: How to preserve privacy of individuals, when learning cumulative  functions of databases, e.g. medical records, financial records? Related work: Crypotography community: strong privacy definitions.  Database,  datamining community: learning algorithms without strong privacy guarantees. Contributions: A logistic regression algorithm that respects strong privacy  definitions, yet has learning guarantees. A new privacypreserving ML technique, applied to logistic regression. A generalization error bound for new algorithm vs. standard logistic regression. Experiments: improvements vs. previous method, and comparison to standard.

Please see poster M54 for details!