Enhancing visual quality of spatial image steganography using SqueezeNet deep learning network

The aims of improving steganographic method are divided into two groups: the first is to make the hiding capacity as high as possible; the second is to make the visible distortion as low as possible. The higher the visual quality of the stego-image, the less suspicious it becomes, which can increase security. However, the distortion caused by embedding data into images is not predictable and typically image dependent. If the user has a database of possible cover images, finding a suitable cover image that can sustain high visual quality after embedding is challenging. Thus, an automatic cover selection method is needed. In this paper, the problem of visual quality of the stego-image is tackled as a classification problem, where a CNN-based classifier is employed to select images that can have high imperceptibility after the process of embedding. To achieve that, a CNN was trained to classify images into “High Quality” and “Low Quality”. The CNN was based on SqueezeNet architecture, and was trained in two scenarios; transfer learning and learning from scratch. The two classifiers were able to achieve very high classification accuracies of F1 = 0.926 and 0.904.