基于隐写噪声深度提取的JPEG图像隐写分析

JPEG image steganalysis based on deep extraction of stego noise

当前基于深度学习的隐写分析方法检测效果受限于其获取的隐写噪声的精确度。为了获取更加准确的隐写噪声,提高隐写分析的准确率,提出了一种基于隐写噪声深度提取的JPEG图像隐写分析方法。首先,设计了隐写噪声深度提取网络,通过有监督的学习方式使网络能够准确地提取载秘图像中包含的隐写噪声;而后,利用设计的模型评价指标选择最优的隐写噪声提取网络;最后,根据隐写噪声的特点设计分类网络,实现载秘图像的检测,并将分类网络与隐写噪声深度提取网络融合得到最终的检测网络。实验在两个大规模的公开数据集(BOSSBase和BOWS2)上进行,针对两种自适应JPEG图像隐写方法(J-UNIWARD和UED-JC)在多个不同嵌入率和图像质量因子条件下构建的载秘图像进行检测。实验结果表明,该方法的检测准确率较性能第二的方法分别提高了约2.22%和0.85%。文中方法通过提取更加准确的隐写噪声,减少了图像内容对隐写分析带来的影响,相比于典型的基于深度学习的JPEG图像隐写分析方法,取得了更好的检测效果。

The performance of steganalysis is limited by the quality of the stego noise obtained by current deep learning-based methods.In order to obtain more accurate stego noise and improve the accuracy of steganalysis,a new method is proposed based on deep extraction of stego noise for JPEG image steganalysis.First,a stego noise deep extraction network is formulated to precisely extract the stego noise from stego images with the supervised trained network.Then,a model evaluation index is proposed to select the most effective network for stego noise extraction.Finally,according to the characteristics of stego noise,a classification network is designed to detect the stego images,which is then combined with the stego noise extraction network to obtain the final detection network.In the steganalysis experiment,two large-scale publicly available datasets(BOSSBase and BOWS2)are used to construct the stego images by two adaptive JPEG image steganography methods(J-UNIWARD and UED-JC)under several embedding rates and quality factors.Experimental results show that the detection accuracy of the method proposed in this article has been improved by up to 2.22%and 0.85%,respectively compared to the second-best performing method.By extracting more accurate stego noise and reducing the impact of image content on steganalysis,the proposed method achieves a better detection performance compared to typical deep learning-based JPEG steganalysis methods.