[Paper Review] JPEG Steganalysis Based on DenseNet

Jianhua Yang, Yun-Qing Shi+2 more|arXiv (Cornell University)|Nov 26, 2017

Advanced Steganography and Watermarking Techniques26 references21 citations

TL;DR

This paper proposes a 32-layer DenseNet-based CNN architecture for JPEG steganalysis that enhances feature reuse through dense connectivity, improving information flow and reducing parameters. It further introduces CNN-SCA-GFR, an ensemble method combining the CNN with conventional SCA-GFR, achieving state-of-the-art detection accuracy with 5.67% and 4.41% lower error rates than XuNet at 0.1 and 0.4 bpnzAC, respectively, while using only 17% of XuNet’s parameters.

ABSTRACT

Different from the conventional deep learning work based on an images content in computer vision, deep steganalysis is an art to detect the secret information embedded in an image via deep learning, pose challenge of detection weak information invisible hidden in a host image thus learning in a very low signal-to-noise (SNR) case. In this paper, we propose a 32- layer convolutional neural Networks (CNNs) in to improve the efficiency of preprocess and reuse the features by concatenating all features from the previous layers with the same feature- map size, thus improve the flow of information and gradient. The shared features and bottleneck layers further improve the feature propagation and reduce the CNN model parameters dramatically. Experimental results on the BOSSbase, BOWS2 and ImageNet datasets have showed that the proposed CNN architecture can improve the performance and enhance the robustness. To further boost the detection accuracy, an ensemble architecture called as CNN-SCA-GFR is proposed, CNN-SCA- GFR is also the first work to combine the CNN architecture and conventional method in the JPEG domain. Experiments show that it can further lower detection errors. Compared with the state-of-the-art method XuNet [1] on BOSSbase, the proposed CNN-SCA-GFR architecture can reduce detection error rate by 5.67% for 0.1 bpnzAC and by 4.41% for 0.4 bpnzAC while the number of training parameters in CNN is only 17% of what used by XuNet. It also decreases the detection errors from the conventional method SCA-GFR by 7.89% for 0.1 bpnzAC and 8.06% for 0.4 bpnzAC, respectively.

Motivation & Objective

• To improve detection of low-SNR steganographic signals in JPEG images using deep learning.
• To enhance feature reuse and information flow in CNNs for steganalysis by leveraging dense connectivity.
• To reduce model complexity and parameters without sacrificing performance.
• To integrate deep learning with conventional JPEG steganalysis methods (SCA-GFR) for improved robustness and accuracy.
• To establish a new SOTA benchmark in JPEG steganalysis using end-to-end deep learning with hybrid architecture.

Proposed method

• The proposed CNN uses 32 layers with dense connections, concatenating features from all previous layers of the same spatial size to enhance feature reuse and gradient flow.
• Preprocessing involves high-pass filtering and truncation of decompressed JPEG images to extract stego-related features.
• Batch normalization and ReLU activation are applied after each convolutional layer to stabilize training and improve convergence.
• The classifier uses a fully connected layer followed by Softmax to output cover/stego probabilities.
• An ensemble method, CNN-SCA-GFR, combines 9 trained CNN models (each outputting 160D features) with SCA-GFR (17,000D features) by concatenating features and averaging final probabilities.
• The ensemble uses intermediate features from the average pooling layer of CNNs, not just the final Softmax output, to improve diversity and accuracy.

Experimental results

Research questions

• RQ1 Can dense connectivity in CNNs improve feature propagation and detection accuracy in low-SNR JPEG steganalysis?
• RQ2 Can a hybrid ensemble of deep learning and conventional steganalysis methods (SCA-GFR) outperform standalone deep learning models?
• RQ3 How does the proposed CNN compare in parameter efficiency and performance to state-of-the-art methods like XuNet?
• RQ4 Does the proposed architecture generalize well across different datasets such as BOSSbase, BOWS2, and ImageNet?
• RQ5 Can the integration of selection-channel-aware features within a CNN framework further enhance steganalysis robustness?

Key findings

• The proposed CNN-SCA-GFR ensemble reduced detection error by 5.67% compared to XuNet at 0.1 bpnzAC and by 4.41% at 0.4 bpnzAC.
• The method decreased error rates by 7.89% compared to conventional SCA-GFR at 0.1 bpnzAC and by 8.06% at 0.4 bpnzAC.
• The proposed CNN used only 17% of the parameters (0.88M) compared to XuNet (5.48M), significantly improving model efficiency.
• On the BOWS2 dataset, the proposed method achieved lower error rates than both XuNet and SCA-GFR across all payload levels, demonstrating robustness.
• On ImageNet, the proposed CNN achieved 15.34% error rate at 0.4 bpnzAC, outperforming XuNet (22.28%) and showing strong scalability on large-scale data.
• The validation error curve on ImageNet showed faster convergence and lower error for the proposed CNN compared to XuNet.