A reproducibility study on: "Random Sampling Plus Fake Data: Multidimensional Frequency Estimates With Local Differential Privacy" by H. Arcolezi, J. Couchot, B. Al Bouna and X. Xiao
Teodora Stamenkovic Luka Kalezic · Zenodo (CERN European Organization for Nuclear Research) · 2023
Local differential privacy (LDP) enables users to protect their data before sending it to the server. Frequently, the server wants to estimate the number of users for each value in multiple attributes (d >= 2). To do this, the privacy budget must be managed carefully due to the composition theorem. Two known solutions are splitting the privacy budget for each attribute (Spl) and random sampling one attribute with all the privacy budget (Smp). Smp has proven to have higher data utility than Spl, but the sampled attribute is visible to the server. The paper we chose proposes a new solution, Random Sampling plus Fake Data (RS+FD), which generates fake data to protect the sampled attribute and uses amplification by sampling. RS+FD is said to achieve the same or better utility than Smp and this claim was supposedly validated using synthetic and real-world data sets. The goal of this paper is to try to reproduce already existing solution, which is previously described, as well as do the significance testing in order to confirm these results.