On the Use of Iterative Feedback in Private Frequency Estimation
Name
Richardson-yaatehr-meng-eecs-2021-thesis.pdf
Description
Thesis PDF
Size
28.06 MB
Format
Adobe PDF
Checksum (MD5)
53c9027592293b2e90e24f68959b83fb
Author(s)
Richardson, Yaateh
Advisor(s)
Raskar, Ramesh
Date Issued
June 2021
Publisher
Massachusetts Institute of Technology
Abstract
This thesis evaluates iterative feedback as a mechanism to optimize privatization algorithms, specifically the family of Local Differential Privacy (LDP) algorithms. The main contribution is the Iterative LDP Algorithm which uses iterative feedback to outperform generic LDP mechanisms in certain scenarios, including those with low response rates (number of samples). This addresses a major pain point of LDP mechanisms currently, as they are prone to significant losses in utility in those scenarios. Conversely, it is less effective than existing methods as the number of responses tends to infinity.
This work contains extensive comparison of the Iterative LDP Algorithm to several existing LDP mechanisms and baselines for frequency estimation. Experiments show gains in estimation performance in high privacy regimes over synthetic and real data. These experiments help discern scenarios where the Iterative LDP Algorithm actually learns from regularization induced performance gains. Learning and substantial performance gains were observed when samples are generated from the power law family of distributions (distributions that look linear on logarithmically scaled axes). Several epidemiological, social network, and other important internet based counting phenomena are known to follow such distributions[25][5]. The Iterative LDP Algorithm outperforms existing variable privacy LDP mechanisms in the aforementioned regimes. Thus, iterative feedback is a viable enhancement for existing LDP mechanisms.
MIT Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
Terms of Use
In Copyright - Educational Use Permitted
Copyright MIT
Persistent DSpace Link
