A Privacy Preserving Onboard Satellite Image Classification Technique Incorporating Homomorphic Encryption and Transfer Learning

Abhijit Roy¹Mahendra Kumar Gourisaria²Rajdeep Chatterjee²Amitkumar V. Jha^2,3*Bhargav Appasani²NICU BIZON^4*Alin Gheorghiță MAZĂRE^4*

• ¹Indian Institute of Information Technology Guwahati, Guwahati, India
• ²Kalinga Institute of Industrial Technology Deemed to be University, Bhubaneswar, India
• ³KIIT University, Bhubaneswar, India
• ⁴Universitatea Politehnica din Bucuresti, Bucharest, Romania

Satellite image classification is a very important and challenging task in the modern technological age. Satellites can capture images of danger-prone areas with very little effort. However, the quantity of satellite images is very high for rapid capturing of images from space, which requires a huge amount of memory to store the data. Besides, keeping the satellite image private is another important task for security purposes. After all, the onboard instant accurate classification with a smaller number of satellite images is a challenging task, which is very important to determine the specific condition of an area for instant monitoring. In the proposed hybrid approach, the captured images are kept secure while the required training of the classification is done separately. Finally, the trained module is encrypted to be used by the satellite to perform the on-board classification task. The Brakerski/ Fan-Vercauteren (BFV)-based homomorphic encryption of EuroSAT satellite images is applied to store images in cloud storage, while the privacy of the images can be kept alongside. Later, the decrypted images are used for training of four transfer learning models (YOLOv8, YOLOv12, ResNet34, ResNet101 and Vision transformer classification). The best-trained module is encoded and encrypted again by using holomorphic encryption to keep the modules limited to authorized devices. The encrypted module is decrypted and decoded to get back the trained module, which is used for the classification of test images instantly. Finally, from the test results, the performance of the transfer learning models is evaluated. The Vision transformer classifier achieved the highest accuracy of 99.65%.