Securing the visual channel: How my car saw the light and stopped learning

Autonomous driving and other applications depend critically on sensor data to function correctly. The most common sensor data is visual data, processed with computer vision algorithms. The effectiveness of these algorithms has markedly improved with the advent of deep learning. Yet, at the same time, research has shown these algorithms bring serious security concerns. Small, even invisible perturbations can completely change the algorithm's interpretation of the data. Basically, we cannot at present trust a car to read a road sign. In this position paper we present the difficulties with reliance on current deep learning algorithms and outline paths forward. In particular, we analyze the security tradeoffs of an alternative approach for visual communication based on visible light communications (VLC).