MixNet: A Robust Mixture of Convolutional Neural Networks as Feature Extractors to Detect Stego Images Created by Content-Adaptive Steganography

Digital steganography, the modern version of the ancient art of data hiding is a prevalent tool for covert communication. Steganalysis, at least as old as steganography comes handy to unearth such hidden channels. The illegal act of information hiding through steganography of digital images can be overcome effectively only by using intelligent steganalytic techniques. In this paper, a novel MixNet framework comprising of six Convolutional Neural Networks (CNNs) is proposed as feature extractors for accomplishing generic steganalysis of spatial content-adaptive algorithms with better detection accuracy. Since the spatial content-adaptive algorithms embed secret bits in the hard to model components of the image like edges or textures, inputs to the CNNs are initially filtered using high pass filters to obtain the embedded content in the form of noise residual. Hierarchical features extracted from these networks are then concatenated and used to train Support Vector Machine classifier. Experimentation is performed using the benchmark BOSSbase v1.01 cover images and stego images are created with three state-of-the-art algorithms HUGO-BD, S-UNIWARD and WOW at five relative payloads 0.1–0.5 bits per pixel (bpp). The experimental results show that the proposed MixNet outperforms the compared related works in literature and proves the robustness of MixNet in detecting content-adaptive steganography.