A generalized image steganalysis approach via decision level fusion of deep models

Recent advancements in digital technologies has greatly facilitated the huge growth of complex images over the internet channels leading to security threats causing unauthorised information access. Such complex images are perceived as a reliable means for secret communication. Hence an active research is carried out for steganalysis – a method to determine the presence of hidden information in the multimedia files. The primary challenge faced by the existing steganalysis approaches is, to be able to extract and learn high level feature representations from images of high texture complexity which is somewhat difficult to be achieved by a single deep learning(DL)-based steganalyzer. In this work, we expound an ensemble of deep models with decision level fusion strategy in predicting an image as cover or stego. The ensemble of validated reference models aims to achieve better performance than using a single CNN architecture. The proposed model comprises of three potent pre-trained deep fine-tuned models- VGG19, ResNet50 and Inceptionv3 followed by a majority voting scheme for spatial image steganalysis task. In comparison to convolution neural networks (CNN) specifically created for steganalysis, such as Qian-Net[22],XU-Net[25],Ye-Net[30],Yedroudj-Net[31], Zhu-Net[29], SR-Net[52], GBRASNet[34], trained from scratch, the suggested model performs significantly better. The proposed framework is compared against eight existing competitive state-of-the-art (SOTA) models over a two class dataset. Experiments are conducted on benchmark image dataset BOSSBase1.01 and BOWS2. This claim is substantiated experimentally on two well known content-adaptive steganographic algorithms WOW and S-UNIWARD with payloads 0.2bpp and 0.4bpp respectively. Extensive experiments and evaluations reveal that our proposed approach yields an accuracy of 99.16% on WOW (0.2bpp), 99.21% on WOW (0.4bpp), 99.07% on S-UNIWARD (0.2bpp), and 99.69% on S-UNIWARD (0.4bpp) steganography algorithms. Moreover, our approach results in better generalization performance, reduced training time and increases accuracy (ACC) over the existing state-of-the-art (SOTA) steganalytic architectures.