Secure and Efficient Image Encryption via LSTM-Enhanced Chaotic Systems

Abstract

This study aims to utilize a chaos-based color image cryptosystem by applying deep learning. In particular, our method uses a four-dimensional chaotic system, whose increased dimensionality enables greater complexity and randomness, both of which are crucial considerations in image encryption. It employs several analytical techniques to ensure the system is chaotic, including the largest Lyapunov exponent and the 0-1 test for chaos. Later, sequences generated from the four-dimensional chaotic system will be used to train a Long Short-Term Memory (LSTM) network. The LSTM Network will predict four new sequences, which need to be tested for chaos verification and randomness to confirm their applicability in encryption. Secure ciphering of the image is performed by applying two proposed scrambling methods and one diffusion method to a chosen color image using the processed chaotic sequences. Due to the inherent complexity of deep learning and improved properties of chaotic signals, decrypting an image without prior knowledge of system parameters is quite difficult for an attacker. Comprehensive security analysis shows that the proposed encryption method achieves optimal results, such as high entropy ( 7.9978), a large key space analysis of 110840, and great differential attack shown by NPCR ( 99.61%) and UACI ( 33.44%). By this method, pixel correlation is successfully broken, offering a strong defence against various cryptographic threats. These results demonstrate that the proposed algorithm, incorporating deep learning and chaos theory, provides a high level of security to meet the significant requirements of robust color image encryption.