StegoFusion-Net: Fusion of Convolutional Neural Networks for Spatial Image Steganalysis

In this paper, we establish a deep learning-based fusion method to enhance the detection accuracy of Stego images. We designed and trained a new convolutional neural network architecture. The proposed CNN model consists of three blocks: a preprocessing block, a feature extraction block, and a classification block. The preprocessing block uses SRM and Gabor filters to extract the steganographic noise. The extracted noise is then passed to the feature extraction block to characterize the stage noise. The classification block uses the extracted features to distinguish between cover and stego images. During the training phase, we divide the input image into four quadrants and train four CNN models, each one quadrant. Similarly, we divided the testing image into four quadrants and predicted the cover/stego probability of each quadrant. We established a fusion technique to combine decisions made by each CNN model into a final decision. We tested the performance of our method using three content-adaptive steganographic algorithms: HUGO, S-UNIWARD, and WOW. We compared our results against the state-of-the-art methods, our method outperformed these methods except in one case. Under the WOW, with 0.4 bpp payload, our method ranked second after GBRAS-Net.