一种用于彩色图像隐写分析的卷积Transformer网络
CTNet: A Convolutional Transformer Network for Color Image Steganalysis
-
摘要:研究背景 图像隐写旨在以不可察觉的方式将秘密信息嵌入数字载体图像以实现隐秘通讯目的。作为博弈的双方,图像隐写分析则试图根据隐写留下的嵌入痕迹来检测含有隐藏信息的隐写图像。由于大多数现代隐写方法主要基于难以建模的图像内容进行嵌入,在一定程度上增加了检测难度。因此,图像隐写分析技术面临着巨大的挑战。目的 当前的图像隐写分析方法大多是针对灰度图像设计,且无法有效地用来检测彩色隐写图像。同时,已有的彩色隐写分析方法都是基于卷积神经网络且其检测准确率仍然没有达到较好的表现。因此,为了有效地提升彩色隐写分析的检测准确率,我们致力于设计一种更有效的隐写分析模型来提取更全面的隐写分析特征。方法 我们提出一种称为CTNet的彩色图像隐写分析网络,该方法包含预处理,特征提取和分类三个模块。预处理模块用来提取输入图像的每个通道的截断残差特征然后将其进行串联,特征提取模块在前期使用卷积结构来提取局部感受野特征,并且在后期采用Transformer结构来获取全局的隐写分析特征,最后的分类模块由一个全局协方差池化层和有着丢弃操作的两个全连接层组成。结果 我们在隐写和隐写分析任务中常用的彩色ALASKA II数据集和灰度BOSSBase和BOWS2数据集上实验。所提出的模型在彩色隐写图像的检测表现均明显优于所对比的工作,同时在灰度隐写图像上也有着可竞争的表现。此外,我们对各个模块都制定了消融实验来验证所提出模型的合理性。最后,我们也在更大数据集和载体源失配等情况下进行了相应的实验来证明模型的鲁棒性。结论 我们提出彩色图像隐写分析网络CTNet,通过合理地设计网络的各个模块来提取更全面有效的隐写分析特征,进一步提升了检测彩色隐写图像的水平。同时,所提出的模型对于检测灰度隐写图像也有着较好的表现。此外,我们的算法框架也可以被应用于彩色JPEG图像隐写分析任务。Abstract: Compared with convolutional neural network (CNN), Transformer can obtain global receptive field features more effectively and has recently achieved great success in natural language processing and computer vision. Due to the particularity of steganography, however, almost all existing steganalytic networks just employ CNN with local receptive fields to detect embedding artifacts. In this paper, we propose a novel convolutional Transformer network for color image steganalysis. Specifically, we firstly obtain various image residuals for each color channel of an input image in the pre-processing module. To capture more comprehensive steganalytic features, the truncated residuals after channel concatenation will pass through a feature extraction module composed of a CNN group and a Transformer group. The CNN group aims to extract local receptive fields features, while the Transformer group with multi-head self-attention as the key tries to extract global steganalytic features. Finally, we employ a global covariance pooling (GCP) and two fully-connected (FC) layers with dropout for classification. Extensive comparative experiments demonstrate that the proposed method can significantly improve the detection performances in color image steganalysis and achieve state-of-the-art results. Although the proposed method is originally designed for color images, it can also obtain competitive results for grayscale images compared with the current best detector. In addition, we provide numerous ablation studies to verify the rationality of the proposed network architecture.