I am interested in the generation of scenarios for the learning and validation of a decision-making function for steering of autonomous vehicles. Scenario data collection driving in real-life situations (e.g. with regard to the number of lanes, surrounding vehicles or weather conditions) is very expensive. One solution is to automatically generate realistic scenarios, respecting the physical constraints of the environment and nearby vehicles, to meet the expectations of autonomous driving and cover as many situations as possible in order to anticipate potential hazards. To do this, I use adverse attack methods thanks to reinforcement learning algorithms. These methods refer to machine learning algorithms that attempt to “deceive” models (for example neural networks) by providing erroneous information. I am thus able, on the basis of existing scenarios, to generate new ones that can mislead the autonomous driver thanks to the various adverse attacks set up, while maximizing the loss function of the decision system taken between the two scenarios.