A Performance-Preserving Scheme for Classifying Differentially Private Remote Sensing Images

Privacy preservation of remote sensing images is often necessary to effectively balance data utilization and sensitive information conservation. Differential privacy (DP) offers a rigorous framework for privacy preservation and it is widely adopted in various domains. However, in the domain of remote sensing, effective privacy-preservation of input images using DP method is mostly achieved with a cost of significantly degraded performance in downstream data mining tasks, such as classification. We address this issue by proposing a performance-preserving DP-based input perturbation approach, termed as P2DP. Typically, P2DP employs a controlled noise-based shape obscurity scheme so as to confound adversaries attempting to decipher the precise details from remote sensing images. Experimentation on benchmark UCMerced and WHU-RS19 datasets validates the efficacy of the proposed P2DP approach. Our research contribution collectively advances the state of the art in privacy-preserving techniques for remote sensing images, offering a practical and customizable solution for balancing privacy concerns with the need for accurate and informative data analysis.