Legislating Autonomous Vehicles against the Backdrop of Adversarial Machine Learning Findings

Recent studies on adversarial machine learning¹ made Michael Grossman, a Texas-based injury lawyer, skeptical of the viability of autonomous vehicles.² These studies had pointed out that adversarial attacks or perturbations on pictures makes it difficult for the algorithm to correctly classify the content of that picture. If this is applied to traffic sign recognition, simple graffiti on the sign could mislead the algorithm that is analyzing the picture of the traffic sign captured by the camera. ³ Rather than recognizing the traffic sign for what it is, the algorithm could attribute a different meaning to the traffic sign. The consequences could be disastrous, especially if, for example, a stop sign would be read as a speeding sign.⁴ When rational car manufacturers know this defect, they will not proceed with the marketing of autonomous vehicles.