FPFnet: Image steganalysis model based on adaptive residual extraction and feature pyramid fusion

Image steganalysis is a technique for detecting images that contain hidden information. Convolutional neural networks have shown great potential in the steganalysis field. In this paper, we propose the feature pyramid fusion network (FPFnet), an image steganalysis model that combines adaptive residual extraction and feature pyramid fusion. In pre-processing, we present an adaptive residual extraction method instead of manually designed filters for extracting diverse residual features. The residual calculation function is reformed to improve the stability of optimal parameters. The residual is adaptively scaled to improve the truncation process. The residual extraction and truncation processes are incorporated into network training. To improve the utilization of residual features in different layers, we design feature pyramid fusion structure by introducing up sampling and feature fusion methods to fuse residual maps of different sizes in neural networks. Comparative experiments with different residual extraction methods, as well as up sampling and feature fusion methods, show that FPFnet has higher accuracy than other steganalysis models on the spatial universal wavelet relative distortion (S-UNIWARD) and wavelet obtained weights (WOW) datasets.