Annihilation of Image Stegogram Through Deep Texture Extraction Based Sterilization

This work presents a deep feature-based image sterilization algorithm that effectively mutilates stego-secrets concealed within loss-less cover images using any steganography scheme without oppressing the perceptual quality of the cover. The proposed method learns the possible steganalytic features such as edges, corners, textures, blobs etc., through a convolution based deep feature extractor network and produces an edge map locating probable stego-secret rich regions of a suspected stego-image. Sterilization by pixel modification is performed by constructing a hexagonal skeleton of the given image using the edge map. The use of deep features and the hexagonal neighbourhood reduces the degradation of visual quality caused by spatial modifications of image pixels. The efficacy of the method is assessed on an experimental image set of size 45,135 containing 5,015 images from USC-SIPI and Boss-Base 1.01 dataset and their secret-embedded versions generated by eight benchmark steganography schemes. The proposed method is observed to produce optimal outcomes in terms of sterilization percentage, sterilization time and visual quality among the relevant state-of-the-art works.