What Can We Learn Privately Shiva Prasad Kasiviswanathan

What Can We Learn Privately? Shiva Prasad Kasiviswanathan, Homin K. Lee, Kobbi Nissim, Sofya Raskhodnikova Adam Smith First published 2008 Presentation by Or Shahaf

ML reminder: problem •

ML reminder: sample •

ML reminder: goal •

Adding Privacy •

Hierarchy of Private Learning Algorithm Classes • Each class is a certain set of classification problems. PAC* = PPAC* • We’ll prove part of this hierarchy. LI* = SQ* LNI* = NASQ*

ML Reminder: PAC learnable •

ML Reminder: agnostically PAC learnable •

PAC* = PPAC* • PAC* = PPAC* LI* = SQ* LNI* = NASQ*

Generic Private Agnostic Learner •

ML reminder: errors •

Generic Private Agnostic Learner: Utility Proof 1 •

Generic Private Agnostic Learner: Utility Proof 2 •

Generic Private Agnostic Learner: Utility Proof 3 •

Generic Private Learner vs. Non-private •

The Local Model • Local Algorithm

The Statistical Query Model • SQ Algorithm

Local model vs. SQ model •

Simulate SQ Algorithm by Local Algorithm SQ Algorithm

Simulate SQ Oracle by Local Randomizers •

Simulate SQ Algorithm by Local Algorithm • Local Algorithm SQ Algorithm

Simulate SQ Algorithm by Local Algorithm •

Local and SQ Learnability •

LI*, SQ* • PAC* = PPAC* LI* = SQ* LNI* = NASQ*

Implications •