Cross-algorithm image steganalysis based on shared representation learning

doi:10.1016/j.knosys.2026.116830

Cross-algorithm image steganalysis based on shared representation learning

2026-08-21

0

PRE

AI

DOI：10.1016/j.knosys.2026.116830

Original

Original request for help

Save

Abstract

En 中文

Current image steganalysis methods achieve high detection accuracy when training and testing use the same embedding algorithm; however, performance deteriorates sharply once the test-time algorithm changes, exposing a severe domain-shift problem. To address this issue, this paper proposes a cross-algorithm steganalysis network designed to enhance model transferability under varying embedding algorithms. First, a source-domain representation learning mechanism is designed, in which the supervised objective explicitly strengthens shared steganographic perturbations across algorithms. Simultaneously, an ImageNet-pretrained ResNet is deeply fused with the residual main branch to improve sensitivity to both texture and semantics. Second, a local adaptation strategy based on trust-region constraints is designed, where a boundary radius is set in the feature space to limit the update magnitude, thereby suppressing overfitting and preserving the stability of shared representations. Finally, under a cross-algorithm experimental setting with the dual exclusion of image sources and image IDs, the results demonstrate that the proposed method achieves better detection performance than several advanced steganalysis networks. Moreover, a series of ablation studies further verifies the effectiveness of the proposed method.

Keywords:

Image steganalysis

Cross-algorithm

Shared representation

Trust-region constraint

Local adaptation strategy

Cross-algorithm image steganalysis based on shared representation learning

Abstract

Journal

Organization