Defensive Distillation Based End-to-end Auto-encoder Communication System

The new generation of wireless communication systems proposes the vision that artificial intelligence should play a more significant role in the development of techniques, thus the application of deep learning (DL) in communications has attracted much attention. DL based auto-encoders have been considered to model an end-to-end communication system. In the situation of DL based communication systems, security problems also need to be concerned. Attackers may utilize the broadcast nature of wireless channels to add perturbations which lead to error transmission. In order to defend against such attacks, a defensive distillation based end-to-end system is proposed. Distillation is introduced to modify the structure of the neural network so that it assists to enhance the robustness of the system model against perturbation and decrease error transmission. Additionally, the robustness against channel noise is also demonstrated in the research.