Securing image data in resource-constrained Internet of Things (IoT) environments necessitates lightweight yet robust cryptographic solutions. This paper presents a novel image encryption model that, for the first time, integrates a session-specific neural network (NN)-generated Substitution Box (S-box) with a chaotic logistic map and the ASCON lightweight authenticated encryption algorithm. The NN, trained with a custom loss function, produces a 16 \(\times \) 16 S-box for each session based on chaotic seeds; a bijectivity correction step ensures that each generated S-box is a true permutation suitable for cryptographic deployment. The same chaotic parameters derive the 128-bit ASCON key, and the neural S-box is em-ployed in ASCON’s substitution layer. Experimental results on 1097 \(\times \) 1166 RGBA images demonstrate strong security metrics: entropy of 7.9999, NPCR of 99.63%, UACI of 33.65%, and an encrypted-image correlation coefficient of 0.1837. Encryption and decryption times are 41.2 s and 44.3 s, respectively. Compared to recent state-of-the-art methods, our approach achieves superior diffusion and randomness at the cost of increased runtime due to dynamic S-box generation. All code, figures, and metrics are available for reproducibility. The proposed framework provides a promising and robust solution for lightweight image encryption in IoT by leveraging the unique strengths of chaotic neural S-boxes within the efficient ASCON algorithm.