Core Concepts
Differentially Private Generative Adversarial Networks (DPGANs) can generate synthetic indoor location data that is statistically similar to the original data while preserving the privacy of individuals.
Abstract
The paper introduces an indoor localization framework that employs DPGANs to generate privacy-preserving indoor location data. The key highlights are:
This is the first work that introduces DPGANs for generating private indoor location data for both location-based and zone-based indoor localization.
The proposed DPGAN framework not only preserves the privacy of indoor location data but also enhances the accuracy of localization.
The paper investigates the influence of two popular DPGANs, Differentially Private Wasserstein GAN (DPWGAN) and Differentially Private Conditional GAN (DPCGAN), on the similarity of the generated datasets to the original dataset, the localization accuracy, and the privacy preservation.
The efficiency and performance of the suggested DPGAN framework for indoor localization are verified through a real-world experimental testbed.
The paper first provides background on indoor fingerprinting localization, differential privacy, and GANs. It then introduces the proposed indoor location DPGAN framework, which consists of training a DPGAN on the original indoor location dataset, generating synthetic data, and using the synthetic data to train a localization model.
The utility evaluation shows that the synthetic data generated by DPWGAN preserves the feature correlations of the original data and achieves similar or better localization accuracy compared to the original data, especially when the number of synthetic samples is increased. The zone-based DPCGAN, on the other hand, shows lower accuracy compared to DPWGAN.
The privacy evaluation demonstrates that the generated synthetic data has a high average Euclidean distance from the original data points, indicating a low disclosure risk and effective privacy preservation.
Stats
The original indoor location dataset contains 384 records with 9 RSS values, (x,y) coordinates, and zone labels.