Differentially Private Heavy Hitter Detection using Federated Analytics

TL;DR

This paper presents practical heuristics and an adaptive hyperparameter tuning method to enhance differentially private heavy hitter detection in federated analytics, focusing on improving accuracy while maintaining privacy and efficiency.

Contribution

It introduces an adaptive hyperparameter tuning algorithm and explores data-selection schemes and deny lists to improve heavy hitter detection under differential privacy in federated settings.

Findings

Improved heavy hitter detection accuracy with adaptive tuning.

Effective data-selection schemes enhance privacy-utility trade-offs.

Demonstrated on Reddit dataset with extensive experiments.

Abstract

In this work, we study practical heuristics to improve the performance of prefix-tree based algorithms for differentially private heavy hitter detection. Our model assumes each user has multiple data points and the goal is to learn as many of the most frequent data points as possible across all users' data with aggregate and local differential privacy. We propose an adaptive hyperparameter tuning algorithm that improves the performance of the algorithm while satisfying computational, communication and privacy constraints. We explore the impact of different data-selection schemes as well as the impact of introducing deny lists during multiple runs of the algorithm. We test these improvements using extensive experimentation on the Reddit dataset~\cite{caldas2018leaf} on the task of learning the most frequent words.