Feature-guided Deep Subdomain Adaptation Network for Dataset Mismatch in Spatial Steganalysis

Abstract

Steganalysis aims to detect covert communication established via steganography. In recent years, numerous deep learning-based image steganalysis methods with high performance have been proposed. However, these methods tend to suffer from distinct performance degradation when cover images in the train and test set are quite different, also known as cover source mismatch. To address this limitation, in this paper, a feature-guided deep subdomain adaptation network is proposed. Initially, the predictions of the pretrained model are used as pseudo labels to divide the unlabeled samples of the target domain into different subdomains, and the distributions of the relevant subdomains are aligned by subdomain adaptation. Afterwards, since the steganalysis model may assign incorrect predictions to samples in the target domain, we integrate guiding features to make the division of subdomains more precise. The experimental results show that the proposed network is significantly better than other three networks such as Steganalysis Residual Network (SRNet), deep adaptive network (J-Net) and Deep Subdomain Adaptation Network (DSAN), when it is used to detect three spatial steganographic algorithms with a wide variety of datasets and payloads. Especially, compared with SRNet, the average accuracy of our method is increased by 5.4% at 0.4bpp and 8.5% at 0.2bpp in the case of dataset mismatch.

Linking Intrinsic Difficulty and Regret to Properties of Multivariate Gaussians in Image Steganalysis

“…The CSM problem in steganalysis is well-known and studied. Most of the research is focusing on mitigating it [3,14,25], while some study the causes of CSM [1,10]. Providing a model for the cover-source is very difficult in practice since it can be possibly defined by infinite and undefined processings, especially ones that are under proprietary software, and/or using non-standard and new algorithms.…”