TechRxiv

Enhanced Privacy in Optical Image Restoration: A Federated Deep Memory-Integrated Neural Network Approach

Download (1.46 MB)
preprint
posted on 2023-12-04, 03:51 authored by Muhammad Jahanzeb KhanMuhammad Jahanzeb Khan, Suman Rath, Muhammad Hassan zaib

In the realm of remote sensing images, restoration and privacy preservation stand as dual challenges. While the intricate characteristics of these images render conventional restoration methods inadequate, concerns regarding data privacy pose a significant barrier to their optimal utilization. Addressing this multifaceted challenge, this study synergizes the Deep Memory Connected Network (DMCN) with federated learning, enabling data-driven model improvements without direct access to the raw image data. This federated approach, while bolstering data privacy, introduces inherent noise into the learning process. To counteract this, techniques such as Gaussian image denoising were employed, ensuring restoration quality. Notably, the federated DMCN exhibited commendable performance, showcasing only a marginal accuracy degradation in the face of noise. Downsampling Units, integral to DMCN, further contributed by reducing computational overheads. Comprehensive evaluations on remote sensing datasets underscore the promise of this federated approach, balancing data privacy with restoration fidelity, and charting a viable path for future applications.

History

Email Address of Submitting Author

jahanzeb@nevada.unr.edu

Submitting Author's Institution

University of Nevada, Reno

Submitting Author's Country

  • United States of America

Usage metrics

    Licence

    Exports

    RefWorks
    BibTeX
    Ref. manager
    Endnote
    DataCite
    NLM
    DC