CRCS Lunch Seminar
Date: Monday, November 5, 2012
Speaker: Kobbi Nissim, Ben-Gurion University and Harvard CRCS
Title: Characterizing the Sample Complexity of Private Learners
Abstract: The notion of private learning [Kasiviswanathan el al. 08] is a combination of PAC (probably approximately correct) learning [Valiant 84] and differential privacy [Dwork et al. 06]. Kasiviswanathan el al. presented a generic construction of private learner for finite concept classes, where the sample complexity depends logarithmically in the size of the concept class. For concept classes of small VC dimension, this sample complexity is significantly larger than what is sufficient for non-private learning.
I will present some of the known bounds on the sample complexity of private learners, and a recent characterization of the sample complexity as a combinatorial measure of the learned concept class.
Joint work with Amos Beimel and Uri Stemmer.