Competitive Coevolution for Color Image Steganography

Abstract

The need for more secure and advanced dissemination is ever increasing since advancement in digital transmission has also posed equally challenging threats. Steganography has been around for a while now, with frequent updates in the traditional methods. In the past one year, with the rise in research in deep learning, Steganography and Steganalysis have emerged as a prime application. The paper has further worked on it, with introducing a competitive coevolution learning approach wherein Generative Adversarial Networks(GANs) are used for performing steganography along with steganalysis to improve the overall results. RGB color images are used as cover and secret images. A mixed loss function was used which is a weighted sum of Mean Squared Error(MSE), Structural Similarity Index(SSIM) and Mean Structural Similarity Index(MSSIM) for generating the stego image and extracting the secret image. The major improvements proposed in this paper are the introduction of attention based Squeeze and Excitation-Inception modules in the hiding and the reveal networks, which enhance spatial encoding and increase the representational power of the network, thereby giving better overall results.