Abstract: Aiming at the problem of extracting more significant steganographic features from images to improve detection accuracy of steganalysis detection, a Siamese network image steganalysis method based on multiattention mechanism is proposed. This method uses the idea of feature fusion to make the steganalysis model extract richer steganographic features. Firstly, a Siamese network subnetwork composed of ParNet block, depthwise separable convolution block, normalizationbased attention module, squeeze and excitation module, and external attention module is designed, and the multibranch network structure and multiattention mechanism are used to extract more useful classification results. Features improve the detection ability of the model; then use Cyclical Focal loss to modify the weight of the training samples at different stages of training to improve the training effect of the model. The experiment uses the BOOSbase 1.01 data set to conduct experiments on five adaptive steganography algorithms: WOW, SUNIWARD, HUGO, MiPOD and HILL. Experimental results show that this method outperforms SRNet, ZhuNet and SiaStegNet methods in detection accuracy, and has a lower number of parameters.

Key words: deep learning, image adaptive steganography, image steganalysis, siamese network, attention mechanism

摘要： 针对从图像中提取更显著的隐写特征来提升隐写分析检测精确度的问题，提出了一种基于多注意力机制的孪生网络图像隐写分析方法.该方法采用特征融合的思想，使隐写分析模型提取更丰富的隐写特征.首先设计由ParNet块、深度可分离卷积块、标准化注意力模块、压缩激励模块、外部注意力模块组成的孪生网络子网，通过多分支网络结构和多注意力机制提取对分类结果更有用的特征提升模型的检测能力；然后使用Cyclical Focal损失在训练的不同阶段修改训练样本的权重提高模型的训练效果.实验使用BOOSbase 1.01数据集，在WOW，SUNIWARD，HUGO，MiPOD和HILL这5种自适应隐写算法中进行了实验.实验结果表明，该方法在检测精度上优于SRNet，ZhuNet和SiaStegNet方法，并且参数量更低.

关键词: 深度学习, 图像自适应隐写术, 图像隐写分析, 孪生网络, 注意力机制