Year: 2024
Type: conference-paper
Abstract: Local Differential Privacy (LDP) was developed as a Differential Privacy (DP) model that protects user data from the collector. However, tasks such as frequency estimation over time are challenging to apply LDP guarantees to, as privacy and utility goals are subjected to increasing privacy budget consumption. Utility can be enhanced through post-processing techniques, but it's important to be aware that they may introduce unintended bias. In this paper, we analyze the performance of a range of longitudinal LDP protocols coupled with various post-processing techniques, of which we determined Norm Sub and PowerNS to be the best-performing and warned against the use of Norm Mul.
Institution: Universidade Federal do Ceará
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4403823909