Screen-Cam Imitation Module for Improving Data Hiding Robustness

Using an attack-simulation module is a well-recognized approach to improving the robustness of end-to-end neural-network-based data-hiding schemes. However, most proposed attack simulators are limited in the types of attacks they cover, usually handling only a basic set of digital transformations. Real, in-demand use cases for data-hiding methods may involve modifications that cannot be modeled by basic digital transformations such as filtering, noise, or compression. In the screen-cam scenario, when an image containing hidden data is displayed on a screen and captured by a camera, the distortions are much more complex and typically require manual experiments that manipulate physical objects in order to replicate. This hinders both the process of creating applicable data-hiding schemes for this scenario and evaluating their effectiveness. In this work, we propose a generator neural network to simulate screen-cam distortions that can replace the manual, time-consuming operations of replicating this attack in the real world, and we show how it can be used to improve the robustness of an existing data-hiding scheme. In our example, we increased robustness by 15% in terms of bit error rate.