Image Deep Steganography Detection based on Knowledge Distillation in Teacher-Student Network

The current hot deep steganography can hide confidential information of an image into a carrier image of the same size. It also has the characteristics of a high steganography rate and is often used for private communication within specific organizations. However, the existing deep steganography detection network is complex and time-consuming. Thus, rapid steganography detection for massive image information is urgently needed. In this study, a typical distillation algorithm is applied to the existing classical deep steganography detection network (Ye-Net and Yedroudj-Net). We constructed a student network, taking Ye-Net and Yedroudj-Net as teacher networks respectively. Appropriate parameters T and α are also selected in accordance with the effect of knowledge distillation. With these parameters, a fast training model of deep steganography detection is established. The proposed deep steganography detection network is compared with Ye-Net and Yedroudj-Net through experiments. Three spatial steganography algorithms (Wow,S-uniward and Hill) and different steganography rates are used. The experimental results show that the proposed network can shorten the training and average detection time by about 70% and 24% when the detection accuracy is slightly lower. Overall, compared with Ye-Net and Yedroudj-Net, the proposed method can achieve fast detection of deep steganography. And code will be available at: https://github.com/hhfshiqi/KD-Ye-net