International Conference on Engineering Technologies (ICENTE'21) , Konya, Turkey, 18 - 20 November 2021, pp.88-92
Image steganography ensures that secret data is hidden in a cover image and aims to transmit the resulting stego image through a communication channel without being noticed by a third party. On the other hand, image steganalysis detects the hidden data in the stego images. Traditional steganalysis techniques focus on obtaining hidden data. However, the presence of secret data must be revealed before obtaining it. Machine Learning (ML) classifiers are used for this purpose with promising high-performance values. ML techniques other than deep learning (DL) require complex and costly feature analysis performed in spatial or transform space. In recent years, DL models have been used to detect the presence of secret messages in the BOSSBase dataset, but there is no study for medical image steganalysis. Therefore, this study aimed to perform medical image steganalysis using a DL model that performs feature analysis on its convolutional layers. An original medical image dataset containing brain MR images was obtained from epileptic patients and healthy volunteers. Two deep convolutional neural networks (CNN) were used. One of them was trained without transfer learning while the feature layers and weights of DenseNet, ResNet, Inception, and Efficient models were transferred to the other one. The training data was obtained by hiding the secret data to the brain images, with different capacity ratios between 0.1 and 1.0 bit per pixel (bpp) using the WOW technique. The results can be summarized in two aspects. First, as expected, the higher the capacity ratio was, the higher classification performance it was obtained. Second, using transfer learning increased the classification performance of the DL model.