EDGAN: Disguising Text as Image using Generative Adversarial Network

摘要

In the concept of data hiding, image is often used as a cover to hide sensitive data inside it. This approach is considered a good addition in securing information to cryptography which only hides the information and not the presence of the message itself. The combination of Deep Learning with Steganography and Cryptography is rarely done. By utilizing Deep Neural Networks to encrypt and hide the messages, it will be increasingly difficult to decrypt and track.In this study, we developed an encryption mechanism to not only conceal messages, but transforming them into images. The image containing the hidden messages can later be decrypted and converted back into the original message. We use Generative Adversarial Network to develop the encryption and decryption models. Text data is converted into a word vector using word2vec model which then used as input for the encryption model to produce the word images. We use the MNIST dataset to train models which are able to produce images that encrypt 1000 word variations. Based on our experiments, we were able to produce robust encrypted images with 98% accuracy of reversible words. We also show that our model is resistant to various minor image attacks such as scaling, noise addition, and image rotation.