Privacy-Enhanced Image Restoration in Remote Sensing via Federated Learning

作     者：Khan, Muhammad Jahanzeb Rath, Suman Zaib, Muhammad Hassan 

作者机构：Dept. of Computer Science and Engineering University of Nevada Reno Reno United States Dept. of Computer Science Air University Islamabad Pakistan 

出 版 物：《TechRxiv》 (TechRxiv)

年 卷 期：2024年

核心收录：

主　　题：Optical remote sensing 

摘      要：This research addresses the dual challenges of image restoration quality and data privacy in optical remote sensing. Traditional restoration methods often fall short due to the complex nature of remote sensing images, and data privacy concerns further complicate the use of advanced techniques. By integrating the Deep Memory Connected Neural Network (DMCN) with the Data-Decoupled Federated Learning (DDFL) framework, our approach enables significant improvements in image restoration without requiring direct access to sensitive raw data. This method not only enhances data privacy by leveraging federated learning principles but also incorporates advanced techniques like Gaussian image denoising to maintain high restoration quality despite potential noise introduced by the federated process. The performance of the federated DMCN, particularly on the UCMERCED dataset, demonstrates minimal accuracy degradation, even in the presence of noise, while the strategic use of Downsampling Units within DMCN optimizes computational efficiency. Our comprehensive evaluations reveal the effectiveness of this approach in balancing data privacy with the need for high-quality image restoration, suggesting a promising direction for future advancements in remote sensing applications. © 2024, CC BY.