A Deep Learning Based Relative Sensitivity Framework for Privacy-Preserving Satellite Image Publishing

Abstract

High resolution satellite images are widely used in many geospatial applications such as urban planning, environmental monitoring, disaster management, agriculture, transportation, and critical infrastructure analysis. But the growing availability and distribution of satellite images can reveal sensitive geographic entities, posing significant privacy and security challenges. In this paper, a Relative Sensitivity-Based Privacy Preservation Framework (RSPPF) is proposed for secure satellite image publishing. The proposed framework introduces a context-aware privacy decision framework that combines deep learning-based geographic attribute identification, sensitivity evaluation, exposure risk assessment and adaptive privacy transformation. To measure the privacy requirement of detected geographic entities, we propose a Relative Sensitivity Score (RSS) that integrates sensitivity level, classification confidence and contextual exposure risk. The framework is designed to select appropriate privacy-preserving transformations such as blurring, pixelation and masking based on the computed RSS values, without degrading the analytical usefulness of non-sensitive regions.
A conceptual demonstration using representative remote sensing scenarios illustrates the operational workflow and decision-making logic of the proposed approach. The proposed framework is intended to enable adaptive protection by considering the importance and risk of individual geographic entities, unlike traditional privacy-preserving methods that provide the same level of protection for all image regions. The framework offers a flexible means of and potential applications balancing privacy preservation and image utility in the dissemination of satellite images. The proposed framework can facilitate future developments in privacy-aware remote sensing systems, with potential applications in geospatial analysis, smart city management, environmental monitoring, and secure geographic information sharing can be found.