Basic Information

• Project Title: Estimation under Label Differential Privacy
• Name: Millen Kanabar
• Project: DDP
• Semester(s): 9, 10
• Guide: Nikhil Karamchandani, Bikash Kumar Dey

Abstract

The aggregate and distribution estimation problems arise naturally in numerous real-world scenarios. While optimal estimators under differential privacy are known for distribution estimation, finding achievable schemes with matching lower bounds for partially private aggregate and mean estimation remains open. In this report, we propose modifications to existing schemes for group-wise aggregate estimation, providing privacy and performance guarantees when parameters are fixed. We also prove a lower bound and a low-privacy-optimal scheme for the more general joint distribution estimation problem for bivariate data, where differential privacy is required with respect to only one of the variates. We show that at under a high privacy requirement, relaxation to partial privacy does not improve the expected risk, whereas only a constant factor improvement can be achieved under low privacy. Finally, in the same setting, we discuss necessary and sufficient privacy conditions when the input distributions are restricted to sets where the two variates are sufficiently uncorrelated, describe candidate schemes for such conditions and provide directions for a possible lower bound.

Any courses you completed relevant to the project

EE 325, 708

Describe your experience on the project

I wanted to work on information theory and privacy; after an SRE project that didn’t go anywhere, I was suggested a paper that I found interesting. I decided to go deeper and found a few improvements and extensions of the problem; that formed the bulk of the project.

Workload was whatever I could do between meetings, and the project was mostly self-driven.

I am working on tying things up before I leave for grad school, if that leads to nice results we might publish them.

Describe your experience with the guide

We had weekly meetings to discuss progress, the guides were very hands off otherwise. We discussed my progress and next steps in every meeting, and I followed up on email if I needed to inform them about the work before the meetings. The guides didn’t delve too deep into the technical details of the work, so most doubts I had were discussed in meetings after I gave them a bunch of background.

Evaluation was super chill. I’d maintained a running progress report that eventually evolved into my final report. I had two presentations with external examiners, one for each stage, but those were very relaxed.

Any advice for anyone considering a project under the same guide? Any other professors working in similar fields?

You’ll have a great time if you are very motivated to do something theoretical. The guides will also try to motivate you to try some simulations if you want results validated. It’s probably a good idea to have company as well (I had friends at my lab that I could discuss technical stuff with, which ended up helping a lot).

Prof. Sibi in EE also does information theory; Prof. Vinod Prabhakaran from TIFR (who also joined as a guide in the second half of the project) does similar work. There might be professors who work on ML and privacy, I am not personally familiar with any myself.