MSCNN: Steganographer Detection Based on Multi-Scale Convolutional Neural Networks

Steganographer detection aims to find guilty user who spread images containing secret information among many innocent users. Feature extraction is an important step in steganographer detection. The challenge is that the existing steganalytic feature extraction of steganographer detection is mostly based on manually designed features. A steganographer detection algorithm based on multi-scale convolutional networks (MSCNN_SD) is proposed to automatically extract steganalytic features in the paper. MSCNN_SD introduces the residual maps and quantization truncation idea of classical SRM into deep convolution network. MSCNN_SD uses a set of filters to extract rich residual information and uses a number of quantized truncation combinations are introduced to discretize the residual maps. Finally, two parallel deep learning subnets are used to learn the features of different scale residuals. The simulation results illustrate that the proposed MSCNN_SD method is superior to the state-of-the-art method. MSCNN_SD uses a well-trained model of one steganography payload, which can detect different steganography and payloads used by steganographer. At the same time, MSCNN_SD has a good detection effect when the steganographer uses the hybrid steganography strategy .