Abstract

In the context of a known plaintext attack, we designed a novel convolutional neural network (CNN) model to evaluate the security of color image encryption techniques. Our model was tested on the two-dimensional improved logistic Coupling map (2D-ILCM). For the training phase, we generated a large set of plaintext and ciphertext pairs from CIFAR10. We managed to efficiently reconstruct the original images by combining the U-Net architecture with dense blocks and an attention mechanism, without the need for encryption keys or internal encryption parameters. This observation is supported by numerical tests showing that the decoded images are visually and statistically similar to the original plain texts with high reconstruction quality.  The accuracy and flexibility of the proposed cryptanalysis model are validated by rigorous quantitative and qualitative evaluations based on multiple cryptographic and similarity metrics. Furthermore, our approach demonstrates its capacity to analyze security flaws in chaos-based image encryption schemes by showing strong resistance to partial occlusion and gaussian noise.

Statistics