AD-RLINN: Adaptive Domain Reversible Lightweight Invertible Neural Network for Secure Image Steganography
Year: 2026
Type: article
Abstract: Multimedia systems increasingly rely on image steganography to secure and transparently transmit information. While several frameworks have been presented for reversible image steganography based on deep learning, they suffer from high computational complexity and instability during adversarial optimization, as well as from insufficient reversibility and inadequate steganalysis resistance. In this work, an adaptive domain reversible lightweight invertible neural network (AD-RLINN) framework for reversible image steganography has been proposed. AD-RLINN encompasses an invertible framework that incorporates adaptive domain-aware embedding, a lightweight reversible attention mechanism, and residual multi-scale feature extraction, as well as adaptive payload allocation and reversible affine coupling. Compared to frameworks that integrate several GANs or transformers, AD-RLINN keeps strict invertibility and reduces computational overhead. A steganalysis-oriented contrastive optimization has also been included in the framework to enhance steganalysis resistance. The results of extensive experiments indicate that AD-RLINN outperforms other approaches (HiDDeN, StegGAN, HiNet, Lossless INN, RISRANet) with regard to PSNR and SSIM, bit error rate (BER), computational time and steganalysis resistance. The approach improves the security of reversible image steganography and decreases the computational cost. This makes it suitable for secure communication in multimedia systems and for other similar intelligent privacy preserving applications.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7225136268