Transferable heterogeneous feature subspace learning for JPEG mismatched steganalysis. (April 2020)

• Record Type: Journal Article
• Authors: Jia, Ju; Zhai, Liming; Ren, Weixiang; Wang, Lina; Ren, Yanzhen; Zhang, Lefei
• Is Part Of: Pattern recognition. Volume 100(2020:Apr.)
• Publication Date: 2020-04
• Subjects: Mismatched steganalysis -- Heterogeneous subspace -- Domain-independent features -- Domain-related features -- Transfer learning; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2019.107105
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: The THFSL method considers the feature space in the source domain and target domain as a combination of domain independent feature and domain related feature subspace, and thus, it can not only inherit the advantages of the existing feature transfer approaches but also combine the unique local and global discrimination information obtained from cover and stego images. We propose a novel distribution discrepancy reduction framework for mismatched steganalysis by linearly reconstructing for target data from source data, which can dramatically reduce the distance between domains. By using sparse representation to model the domain related features, the new proposed method can avoid a potentially negative transfer so that it is more robust to different domain changes in mismatched steganalysis. Abstract: Steganalysis is a technique that detects the presence of secret information in multimedia data. Many steganalysis algorithms have been proposed with high detection accuracy; however, the difference in statistical distribution between training and testing sets can cause mismatch problems, which will degrade the performance of traditional steganalysis algorithms. To solve this problem, we propose a transferable heterogeneous feature subspace learning (THFSL) algorithm for JPEG mismatched steganalysis. Our approach considers the feature space in each domain as a combination of the domain-independent features and the domain-related features. We use the transformation matrix to transfer both the domain-independent and domain-related features from the source and target domains to a common feature subspace, where each target sample can be better represented by a combination of source samples. By imposing low-rank constraints on the domain-independent features, the structures of data can be preserved, which can capture the intrinsic structures for discriminating cover and stego images. Our method can avoid a potentially negative transfer by using a sparse matrix to model the domain-related features and, thus, is more robust to different domain changes in mismatched steganalysis. Extensive experiments on various mismatched steganalysis tasks show the superiority of the proposed method over the state-of-the art methods.