Autoencoder-based image encryption using hybrid scrambling, diffusion, and dimensionality reduction
S Rithesh Manikandan, Linkkesh A V, Sreenivasan S, V Thanikaiselvan, S Subashanthini, Rengarajan Amirtharajan · Results in Engineering · 2026
• A novel Computational Auto Encoder model to compress 256 × 256 grayscale images to 128 × 128, reducing data by 4 times . • A novel CNN-based unique vector generating algorithm based on plaintext image properties. • Chaotic map based pseudorandom sequence generation with Ikeda and Henon Maps using generated vectors. The process results in a key that is robust against security attacks. • An encryption algorithm based on an index, permutation-based confusions and diffusions with pixel-manipulated vectors. • Comprehensive security analysis demonstrates the resistance against common security attacks, ensuring the algorithm’s reliability and protection. • Reconstruction of decrypted and compressed features results in good visual quality with high PSNR. Even in today’s advancing world, there are still areas where multimedia channel security is compromised for various reasons. Encrypting data in its smaller dimensions ensures the effective utilisation of channel bandwidth, saving computational costs and improving performance. To achieve these parameters, a compression-encryption algorithm using Autoencoders, Henon and Ikeda maps using a multi-stage confusion and diffusion process with a randomised transformation is presented in this work. The dimensionality of the plaintext, with a size of 256 × 256, is reduced to 128 × 128 and encrypted using the generated key via a Convolutional Neural Network-based algorithm. The generated cipher in this novel methodology yields an average NPCR and UACI value of 99.649387% and 33.651190%, respectively, surpassing various statistical, differential, and noise attacks in testing and comparison against proposed symmetric methods. The decrypted and reconstructed image is highly similar to its plaintext, yielding an average PSNR value of 33.72963, which is a benchmark critical value for wireless transmissions. This work effectively utilises channel bandwidth by securing and transmitting 1/4 th of the data to retrieve it at the receiver.