STEGANALYSIS OF DATA HIDDEN IN RGB IMAGES USING RESIDUAL NEURAL NETWORK

Digital forensic analysis aims to apply scientific and statistical techniques to identify, gather, preserve and present relevant digital evidence which will allow to confirm or reject a hypothesis against possible criminal activity. Current methods of forensic digital analysis are effective for visual analysis of physical evidence, but they do not allow for the automatic execution of large amounts of data, for correlation studies of the files obtained, for the validation of the metadata and for the identification of abnormalities in text, graphic or audiovisuals files. For this reason, artificial intelligence techniques were proposed for processing data, for identifying patterns and trends that make it possible to perceive aspects that cannot be visually perceived. In this paper, we propose a network analysis approach to steganalysis of RGB images in frequency domain based on Residual Neural Network-50 (ResNet-50). We used the Discrete Cosine Transform (DCT) features based on standard convolutional operations in ResNet-50 generated by the first convolutional layer in the network. The DCT Feature based ResNet-50 model extensively reduces the quantity of parameters and multiplications while keeping comparable accuracy results to normal ResNet in ImageNet-1K.