Towards Explaining AI in Future Communication Systems: Trust & Safety

Deep learning is being used in a wide range of communication functions from channel optimization to radio resource allocation. Unlike traditional model based optimization, we have very weak understanding of the output distributions and hence how extreme values and adversarial attacks can lead to malfunctions. This is increasingly important when we couple communication functions with safety- and mission-critical systems. Here, we examine our work in interpreting and explaining AI actions in a range of supervised and reinforcement learning modules for communications. We also discuss the notions of measuring physical and emotional trust, and how to integrate these practices into the communication system.