Image steganalysis via squeeze-and-excitation and separable convolutional neural network

Konferenz: AIIPCC 2021 - The Second International Conference on Artificial Intelligence, Information Processing and Cloud Computing
26.06.2021 - 28.06.2021 in Hangzhou, China

Tagungsband: AIIPCC 2021

Seiten: 4Sprache: EnglischTyp: PDF

Li, Jingtai; Wang, Xiaodan; Song, Yafei; Wang, Shuo (Air and Missile Defence College, Air Force Engineering University, Shaanxi, China)

For steganalysis, recent research has shown the immense potential of convolutional neural network. With the different architecture apparence, the accuracy and feasibility of steganalysis model are promote to a new stage, but there are leaving a lot room for improvement. Applying novel method from computer version filed is a general approach to improve model preformance. Squeeze-and-Excitation method and Separable Convolution were proved to be an effective structure. Combining with two method, a novel image steganalysis model was proposed to reach a better result. After comparing with different structure of block, having an residual learning layer after feature extraction proved to accelerate convergence of model. The experiments presented that the accuracy of model is similar with previous model, but have less number of epoch to decrese loss.