NGEMD & ADV-NGEMD: a unified framework for high-capacity, efficient, and adversarially secure image steganography

Abstract

Exploring image steganography, one prominent technique is Exploiting Modification Direction (EMD), which is favored for its high efficiency achieved through minimal image alterations. However, this efficiency comes at the cost of low capacity, prompting the development of numerous EMD‐based methods primarily focused on increasing payload. Yet, none have managed to simultaneously deliver both high capacity and optimal efficiency. To address these shortcomings, we introduce NGEMD—a next-generation EMD‐based steganographic framework that determines optimal extraction coefficients and bases solely from the pixel group length, maximum per‐pixel change, and number of modifiable pixels $(n,z,k)$. By deriving recursive relations to establish an ary‐notational system and employing a systematic solution based on the Chinese Remainder Theorem, NGEMD maximizes both capacity and efficiency while significantly reducing computational cost. Due to the inherent weakness of conventional EMD‐based methods against modern steganalysis, we further develop ADV‐NGEMD (Adversarially-NGEMD). We present a scheme to resist deep learning‐based steganalyzers such as YeNet, called ADV‐NGEMD, by considering the hidden message as an adversarial vector and applying changes based on the opposite sign of the gradient while controlling the modifications through a customized cost function. Comprehensive experiments confirm that both NGEMD and ADV‐NGEMD deliver exceptional performance, achieving high payload capacities (up to 2.5 bpp) while preserving visual quality (with PSNR values up to 58 dB and SSIM above 0.99) and, for instance, significantly increasing miss detection rates—from 4 % in NGEMD to as high as 60 % in ADV‐NGEMD at comparable capacities—without sacrificing their high‐capacity advantages.