Deep Residual Neural Networks with Attention Mechanism for Spatial Image Steganalysis
Yue Shu, Renchao Qin, Yaying He, Ya Li, Ruilin Jiang, Zhiyuan Wu · 2022
Image steganalysis has been explored for decades to detect whether an image has hidden secret data. Many recent works have shown that CNNs (Convolutional Neural Networks) trained with rich features perform better than traditional two-step machine learning approaches. Some CNNs reach high precision in the classification task of steganalysis. However, such precision is insufficient for actual security requirements, especially when the image steganography payload is small. To further improve the precision for steganalysis, we propose a residual based CNN model with channel attention mechanism and an innovative pre-processing activation strategy that outperforms previous works in classification precision. Our CNN adopts SRM (Spatial Rich Models) to get rich features and utilizes depth wise separable convolution and Squeeze-and-Excitation attention module to build a CNN with residual structure. In addition, there are two linear fully-connected layers at the end as classifiers. We test our model with WOW and S-UNIWARD steganography at 0.2bbp and 0.4bbp payloads on BOSSbase 1.01 dataset. Our model out-performs previous works, including GBRAS-Net, YEDROUDJ-Net, ZHU-Net, and five others, in terms of classification precision with these four configurations, showing an even more significant precision advantage in the small payload. With the challenging configuration of S-UNIWARD steganography at 0.2bbp payload, the precision of our model is 5.7% higher than the previous best work.