Performance analysis of stego image calibration with the usage of denoising autoencoders

Abstract

Methods for early detection of sensitive information leakage by data transmission in open (public) communication systems have been of special interest. Reliable detection of modified (stego) cover files, like digital images, requires usage of computation-intensive methods of statistical steganalysis, namely covering rich models and deep convolutional neural networks. Necessity of finetuning parameters of such methods to minimize detection accuracy for each embedding methods has made fast retrain of stegdetectors in real cases impossible. Therefore, development of low-complexity methods for detection of weak alterations of cover image parameters under limited prior information about used embedding methods has been required. For solving this task, we have proposed to use special architectures of artificial neural networks, such as denoising autoencoder. Ability of such networks to estimate parameters of original (cover) image from the noisy ones under limited prior information about introduced alterations has made them an attractive alternative to state-of-the-art solutions. The results of performance evaluation for shallow denoising autoencoders showed increasing of detection accuracy (up to 0.1 for Matthews correlation coefficient) in comparison with the state-of-the-art stegdetectors by preserving low-computation complexity of network retraining.