Performance Analysis of Stego Images Detection Using Shallow Denoising Autoencoders

Early detection of sensitive information leakage in communication systems is topical task today. This requires usage of advanced statistical processing methods for detection negligible changes of cover files, such as digital images, caused by message hiding. One of promising approaches for solving the task is learning an appropriate representation of cover and formed stego images that is sensitive to data embedding. This approach is widely used in modern stegdetectors based on utilization of convolutional neural networks. Achieving of high detection accuracy by stegdetector requires usage deep convolutional networks, whose computation-intensive re-train procedure limits fast adaptation to unknown embedding methods. For overcoming this limitation, we propose to use special types of neural networks, namely autoencoders that provides fast adaptation to changes of inputted data by preserving high restoration accuracy. The work is devoted to performance analysis of usage shallow denoising autoencoders for detection of stego images formed by advanced embedding methods. It is revealed that considered networks allows improving detection accuracy up to 1.5%-2% for the most difficult case of small cover image payload (less than 10%).