Efficientnets Aplicadas à Esteganálise Em Imagens Digitais -

Abstract

Several CNN architectures with a specific purpose for steganalysis were developed and reached the state-of-the-art, surpassing the previous models that were based on the feature extraction and classification steps. New image datasets were proposed, differing from the previous ones by the number of instances and the variation of important characteristics such as the quality factor and the payload of hidden messages between images. In addition, new general-purpose architectures are applicable in the scope of steganalysis and benefit from transfer learning to accelerate training. This work presents the training of  the Seteganalys Residual Network (SRNET) with random initialization of weights and performs the performance comparison between the CNN architectures Efficientnet and Efficientnetv2, with the latter perfoming 32% faster than EfficientnetB4, for each training epoch. Finally, an experiment involving successive training within the cover image and its respective stego-images.