A Hybrid Privacy-Preserving Deep Learning Approach for Object Classification in Very High-Resolution Satellite Images

Deep learning (DL) has shown outstanding performances in many fields, including remote sensing (RS). DL is turning into an essential tool for the RS research community. Recently, cloud platforms have been developed to provide access large-scale computing capacity, consequently permitting usage of architectures as a service. However, this opened door new challenges associated with privacy and security data. The data used train algorithms several requirements. Some them need high level confidentiality, such satellite images related public spatial resolutions. Moreover, are usually protected by copyright, owner may strictly refuse share them. Therefore, privacy-preserving deep (PPDL) techniques possible solution problem. PPDL enables training on encrypted without revealing original plaintext. This study proposes hybrid approach object classification very-high-resolution images. proposed encryption scheme combines Paillier homomorphic (PHE) somewhat (SHE). combination aims enhance while ensuring good runtime accuracy. method encrypt maintained through keys PHE SHE. Experiments were conducted real-world high-resolution acquired using SPOT6 SPOT7 satellites. Four different CNN considered, namely ResNet50, InceptionV3, DenseNet169, MobileNetV2. results showed that loss accuracy after applying algorithm ranges from 2% 3.5%, best validation dataset reaching 92%.