Secure Reversible Transformations through Neural Coupling Architecture

This work presents a neural-based cryptographic framework that integrates properties of both block ciphers and stream ciphers through an invertible coupling network. The model employs afixed-key, 128-bit transformation in which encryption and decryption are learned jointly, while an adversarial network attempts unauthorized plaintext recovery. Using Real-NVP-style affine coupling layers, the system ensures exact invertibility and secure reversible transformations. Adversarial training enables near-perfect reconstruction for the legitimate receiver while maintaining high uncertainty for the adversary. By combining fixed transformations with continuous ciphertext outputs and noise-based perturbations, the framework exhibits dual characteristics of classical block and stream ciphers. Experimental results demonstrate the effectiveness of the approach as a hybrid neural cryptographic mechanism, providing secure, learnable encryption without hand-designed structures.