Steganalysis of Context-Aware Image Steganography Techniques Using Convolutional Neural Network

Abstract

Image steganography is applied to hide some secret information. Occasionally, steganography is used for malicious purposes to hide inappropriate information. In this paper, a new deep neural network was proposed to detect context-aware steganography techniques. In the proposed scheme, a high-boost filter was applied to alleviate the high-frequency while retaining the low-frequency details. The high-boost image was processed by thirty SRM high-pass filters to obtain thirty high-boost SRM filtered images. In the proposed CNN, two skip connections were used to collect information from multiple connections simultaneously. A clipped ReLU layer was considered in spite of the general ReLU layer. In constructing the CNN, a bottleneck approach was followed for an effective convolution. Only a single global average pooling layer was used to retain the complete flow of information. SVM was utilized instead of the softmax classifier to improve the detection accuracy. In the experimental results, the proposed technique was better than the existing techniques in terms of the detection accuracy and computational cost. The proposed scheme was verified on BOWS2 and BOSSBase datasets for the HILL, S-UNIWARD, and WOW context-aware steganography algorithms.

An Ensemble Transfer Learning Model for Detecting Stego Images

“…Compared to state-of-the-art methods, the proposed method performed better in terms of detection accuracy and computational cost in the experiments. The HILL, S-UNIWARD, and WOW context-aware steganography algorithms were tested on the BOWS2 and BOSS base datasets, validating the suggested scheme [15].…”

“…In [15], the authors presented a novel deep neural network for identifying contextual steganography methods. The suggested method employed a high-boost filter to reduce high-frequency noise while keeping low-frequency information intact.…”

High-Pass-Kernel-Driven Content-Adaptive Image Steganalysis Using Deep Learning

“…An image BOWS2 dataset [6] is considered in the first row of Figure 1 to study the behavior of the schemes. Residual arrays are created by taking the difference between cover and stego-images [7]. It is apparent from the residual arrays that the number of modified elements is also amplified as the payload increases.…”