An efficient recurrent neural network based confusion component construction and its application in protection of saliency in digital information
Lal Said, Majid Iqbal Khan, Danial Amin · Nonlinear Engineering · 2026
Abstract The explosive growth of Internet of Things (IoT)-driven imaging in medicine, city surveillance, and intelligent infrastructure requires secure, timely transportation of delicate visual information with salient information, like faces and diagnostically important medical areas. Standard block ciphers, including AES, are unable to consistently retain these attributes under burst errors, partial data corruption, or focused cropping. In this paper, we introduce a lightweight substitution–permutation network (SPN) oriented encryption paradigm purpose-built for salient information secrecy in resource-limited IoT applications. We integrate permutation-driven block shuffle by chaos, recurrent neural network (RNN)-guided nonlinear static S-box generation, and bit-parity scrambling at the bit level to improve confusion–diffusion properties. We demonstrate experimental results of NPCR > 99.60 %, UACI > 33.40 %, near-zero correlation, and satisfactory key sensitivity. The technique maintains integrity of the salient region even with 50 % pixel loss, with throughput acceptable for real-time applications. Compared with previous work on lightweight approaches, we provide improved salient feature retention and lower computational complexity, and thus an ideal solution to security-critical applications of IoT-driven imaging.