Mikołaj Płachta
supervisor: Artur Janicki
Information security is becoming an increasingly important issue in our world. As a result, more and more sophisticated attack methods are emerging, and among them are methods using image steganography. Therefore, in the first part of my research for my dissertation, I focused on methods for automatic detection of this type of steganography. The main part of my research is the use of deep machine learning methods. For feature extraction methods, several methods were tested. The first idea was to analyze the discrete cosine transform and teaching a neural network to classify JPEG images. This method did not give satisfactory results, so the next step was to use other coefficients that can extract relevant information from images and help perform detection. For this purpose, DCT Residuals (DCTR), Gabor Filter Residuals (GFR) and PHase AwaRe pRojection Model (PHARM) were used. The BossBase kit, well known in the world of steganography, was used for testing. It consists of carefully selected black and white images in amount of 10,000. Based on it, files with encrypted data were generated using jUniward, nsF5 and UERD algorithms for two data densities of 0.4 and 0.1. Pairs of images (without steganography and with hidden data) were used to train and verify the network. Research was then performed to find a single model that can perform well in multi-threat detection. It was possible to achieve a detection efficiency of 72 percent with a single model.