Abstract
Abstract: Content-adaptive image steganography based on minimizing the additive distortion function and Generative Ad-versarial Networks (GAN) is a promising trend. This approach can quickly generate an embedding probability map and has a higher security performance than hand-crafted methods. however, existing works have ignored the semantic information between neighbouring pixels and the NaN-loss scenarios, which leads to improper convergence. Such cases will degrade the generated Stego images’ quality, decreasing the secret payload’s security. FT GAN performance, which incorporates feature reuse in generator architecture, has been investigated by proposing the FC DenseNet-based generator herein. This investigation explores the superior semantic segmentation capabilities of FC DenseNet, including feature reuse, implicit deep supervision, and the vanishing gradient problem alleviation of DenseNet, toward enhancing visual results, increasing security performance, and accelerating training. The ability to maintain high-quality visual characteristics and robust security even in resource-constrained environments, such as Internet of Things (IoT) contexts, demonstrates the practical benefits of this approach. The qualitative analysis of the visual results regarding the texture regions’ localization and intensity exhibited augmented visual quality. Moreover, an improvement in the security attribute of 0.66% has also been demonstrated regarding average detection errors made by the SRM EC Steganalyzer across all target payloads.