A novel steganalysis method with deep learning for different texture complexity images

Most steganalysis methods based on deep learning don’t distinguish the image texture complexity, but rather to mix all images for training. As a result, the differences of image content between images with different texture complexities are larger than the differences caused by steganographic signal...

Full description

Saved in:
Bibliographic Details
Published in:Multimedia tools and applications 2019-04, Vol.78 (7), p.8017-8039
Main Authors: Zhong, Shangping, Jia, Cunmin, Chen, Kaizhi, Dai, Peng
Format: Article
Language:English
Subjects:
Complexity
Computer Communication Networks
Computer Science
Data Structures and Information Theory
Deep learning
Feature extraction
Image detection
Methods
Multimedia Information Systems
Special Purpose and Application-Based Systems
Steganography
Texture
Training
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Most steganalysis methods based on deep learning don’t distinguish the image texture complexity, but rather to mix all images for training. As a result, the differences of image content between images with different texture complexities are larger than the differences caused by steganographic signal, which is unfavourable for extracting the effective stegananalysis features. In order to reduce the influence of image content difference on steganalysis performance, this paper propose a steganalyais framework adopting the method of training model separately on dividing data sets. Firstly, according to the image texture feature, some correlated statistical features of gray level co-occurrence matrix are used as a measure index of texture complexity, and the data set is divided into several subsets of different complexity according to the index. Secondly, for images with different texture complexities, the corresponding steganalysis models are constructed and trained by using Most Effective Region (MER) and Inception ideas, respectively. For an image to be detected, its texture complexity is calculated first and the corresponding model is used for detection. Finally, an ensemble learning method is used to further improve the framework precision. The experimental results show that our proposed steganalysis method outperforms the handcrafted-features-based steganalysis methods and several CNN-based methods in detection accuracy.
ISSN:1380-7501
1573-7721
DOI:10.1007/s11042-018-6573-5