TY - JOUR
T1 - JPEG steganalysis with combined dense connected CNNs and SCA-GFR
AU - Yang, Jianhua
AU - Kang, Xiangui
AU - Wong, Edward K.
AU - Shi, Yun Qing
N1 - Funding Information:
This work was supported by NSFC (Grant Nos. U1536204, NSFC 61772571) and special funding for basic scientific research from Sun Yat-sen University (Grant No.6177060230).
Funding Information:
Acknowledgments This work was supported by NSFC (Grant Nos. U1536204, NSFC 61772571) and special funding for basic scientific research from Sun Yat-sen University (Grant No.6177060230).
Publisher Copyright:
© 2018, Springer Science+Business Media, LLC, part of Springer Nature.
PY - 2019/4/1
Y1 - 2019/4/1
N2 - The detection of weakly hidden information in a JPEG compressed image is challenging. In this paper, we propose a 32-layer convolutional neural network (CNN) involving feature reuse by concatenating all features from previous layers. The proposed method can improve the flow of gradient, and the sharing of features and bottleneck layers can also dramatically reduce the number of parameters in the proposed CNN model. To further improve the detection accuracy and combine the directional features from the selection-channel-aware Gabor filtering residual (SCA-GFR) method with Gabor filtering and non-directional feature maps from the CNN model, an ensemble architecture called CNN-SCA-GFR is used, which combines the proposed CNN method with the conventional SCA-GFR method to detect J-UNIWARD and UERD. This can significantly reduce the detection error rate to below that of the existing JPEG steganalysis methods. For example, in the detection of J-UNIWARD at 0.1 bpnzAC, the detection error rate using our proposed method is 5.67% lower than that achieved by XuNet, and 7.89% lower than that achieved by the conventional SCA-GFR method. When detecting UERD at 0.1 bpnzAC, the detection error rate using our proposed method is 5.94% lower than that achieved by XuNet, and 10.28% lower than that achieved by the conventional SCA-GFR method.
AB - The detection of weakly hidden information in a JPEG compressed image is challenging. In this paper, we propose a 32-layer convolutional neural network (CNN) involving feature reuse by concatenating all features from previous layers. The proposed method can improve the flow of gradient, and the sharing of features and bottleneck layers can also dramatically reduce the number of parameters in the proposed CNN model. To further improve the detection accuracy and combine the directional features from the selection-channel-aware Gabor filtering residual (SCA-GFR) method with Gabor filtering and non-directional feature maps from the CNN model, an ensemble architecture called CNN-SCA-GFR is used, which combines the proposed CNN method with the conventional SCA-GFR method to detect J-UNIWARD and UERD. This can significantly reduce the detection error rate to below that of the existing JPEG steganalysis methods. For example, in the detection of J-UNIWARD at 0.1 bpnzAC, the detection error rate using our proposed method is 5.67% lower than that achieved by XuNet, and 7.89% lower than that achieved by the conventional SCA-GFR method. When detecting UERD at 0.1 bpnzAC, the detection error rate using our proposed method is 5.94% lower than that achieved by XuNet, and 10.28% lower than that achieved by the conventional SCA-GFR method.
KW - Adaptive steganography
KW - Convolutional neural networks
KW - Ensemble
KW - JPEG image steganalysis
UR - http://www.scopus.com/inward/record.url?scp=85058620795&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85058620795&partnerID=8YFLogxK
U2 - 10.1007/s11042-018-6878-4
DO - 10.1007/s11042-018-6878-4
M3 - Article
AN - SCOPUS:85058620795
SN - 1380-7501
VL - 78
SP - 8481
EP - 8495
JO - Multimedia Tools and Applications
JF - Multimedia Tools and Applications
IS - 7
ER -