Abstract
The main goal of this paper is to investigate the applicability of a back-propagation artificial neural network on the encryption of huge-sized satellite images. The central contribution is using fixed, arbitrary keys in the training process as in classical symmetric and asymmetric cryptography. The used network is of NxMxN neurons representing the input, hidden, and output layers, respectively. The network is trained by adjusting the weights while the bias is given a constant value between 0 and 1 after normalizing the values. A bias is determined. The bias between the input layer and the hidden layer works as the first key (K-1), while the bias between the hidden layer and the output layer represents a second key (K-2).
The training method uses K-1, K2, or both and is done using images of small sizes to improve speed. Then, the network is used to encrypt and decrypt normal satellite images. Numerous trials were done for different satellite optical and SAR images and the goodness of fit (quality of decryption) between the original images and the decrypted ones was at least 98%, even for the images that the network was not previously trained to decrypt. It was also found that the network is not affected by geometrical image distortions like translation, size, and rotation.