Online Program

Encode, Shuffle, Analyze Revisited: Strong Privacy Despite High Epsilon (308173)

*Abhradeep Guha Thakurta, Google Research Brain Team and UC Santa Cruz 

In this talk, I will focus on designing differentially private algorithms that operate in the high-epsilon regime (i.e., epsilon >> 1). Through ideas borrowed from the Encode-Shuffle-Analyze (ESA) framework by Bittau et al., I will show that one can operate with high-epsilon under local differential privacy, and still get very strong central differential privacy guarantees. Furthermore, for a few of the problem settings, the accuracy/privacy trade-offs are optimal under both central and local differential privacy.