Effects and Solutions of Cover-Source Mismatch in Image Steganalysis - Archive ouverte HAL Access content directly
Journal Articles Signal Processing: Image Communication Year : 2020

Effects and Solutions of Cover-Source Mismatch in Image Steganalysis

(1) , (1) , (2) , (3, 4)
1
2
3
4

Abstract

The Cover-Source Mismatch (CSM) has been long recognized as a major problem in modern steganography and steganalysis. Indeed, while a vast majority of works in steganography and steganalysis had been tailored to a specific reference database, namely BOSSbase, recent works show that, because of CSM, the results may greatly differ when changing this dataset. Although the CSM has already been the subject of several publications, these prior works investigated only a few elements in a limited setup. The goal of the current paper is to study the effects of the CSM in a more comprehensive manner and then to examine and compare different strategies for mitigating it. It first defines two different parameters, the source difficulty and the source inconsistency, which are involved in the CSM. Then, using different steganographic schemes and feature sets, it aims at providing a systematic study regarding the various factors that can give birth to CSM for image steganalysis. Finally, two practical ways to mitigate the CSM, using training techniques promoting either diversity of different sources or the specificity of one targeted source which is beforehand identified by training a multi-class classifier, are presented and their performances are compared for different training set sizes.
Fichier principal
Vignette du fichier
ImageCommunication_Final.pdf (1.25 Mo) Télécharger le fichier
Origin : Files produced by the author(s)
Loading...

Dates and versions

hal-02631559 , version 1 (27-05-2020)

Identifiers

Cite

Quentin Giboulot, Rémi Cogranne, Dirk Borghys, Patrick Bas. Effects and Solutions of Cover-Source Mismatch in Image Steganalysis. Signal Processing: Image Communication, 2020, 86, ⟨10.1016/j.image.2020.115888⟩. ⟨hal-02631559⟩
250 View
386 Download

Altmetric

Share

Gmail Facebook Twitter LinkedIn More