Image-Dependent Feature Conditioning for Chaotic Image Encryption

The long-standing limitation of traditional chaos-based image encryption is that the system parameters are fixed, and the cryptographic key is obtained from them, which means that all encrypted images with the same system parameters would be vulnerable to a single known-plaintext or chosen-plaintext attack. This weakness is discussed in this paper, by the use of image-dependent feature conditioning: a mechanism, which directly links the encryption key to the statistical and structural property of each of the plaintext images. Four complementary image descriptors mean intensity, standard deviation, Shannon entropy, and Sobel edge energy are rolled together into a small fingerprint vector F and processed by a three-layer neural network with a small fixed parameter set into a full-resolution spatial modulation map. This map conditions a pseudo-random keystream to a per-image visual key such that any change in pixel content changes the key in every spatial location. The encryption process is then based on a 2-stage confusion diffusion architecture: using image conditioned chaotic permutation provides confusion and a CBC mode XOR cascade using a SHA-256-derived initialisation vector provides ciphertext avalanche (NPCR ≈ 99.6%).