The expeditious advancement of technology and the proliferation of connected devices have created a growing demand for cryptographic algorithms that balance security strength with implementation efficiency. Lightweight cryptography (LWC) aims to address this need, especially for constrained platforms such as IoT and edge devices. While many existing lightweight ciphers prioritize performance, they often fall short against advanced cryptanalytic techniques. This research investigates the integration of neural networks into the substitution layer (S-box) of the CURUPIRA-1 block cipher to improve its cryptographic strength. The goal is to enhance resistance to differential and linear attacks while retaining the cipher’s modular and lightweight properties. By refining the S-box using neural-guided optimization strategies, the proposed method introduces cryptographic adaptability and dynamic instance-specific generation, which may support future deployment in contexts like IoT without compromising core security properties.
