Steganography in the printed world

Abstract

Hiding messages in the digital domain is relatively simple. How to add hidden message in the printed content? Recently deep networks have shown that they are able to detect changes in the printed material which the naked eye can not spot. However the challenge with the printed domain is the increased noise associated with the printing process. In this work we plan to use the deep learning to detect changes in the printed material. This application can be used for enhancing private communication even under censorship. On the other hand the insights from this work might enable us to detect malicious messages that might me in use by technical savvy criminals or terrorists. This study introduces a novel Noise adding approach to conceal messages and content in the images. As we know machine learning algorithms are susceptible to adversarial perturbations we can use this disadvantage to our advantage and send messages across .So in this study we make use of Deep Neural Network disadvantage and convey messages by adding Imperceptible and perceptible(Either Patches or Camouflage) to Character and Classify it with the help or our Character classifier to measure robustness of our process in the printed domain and find the factor that effect these results. The dataset user for this classifier is The Chars74K dataset Character Recognition in Natural Images (characters from computer fonts with 4 variations (combinations of italic, bold and normal)