Robust Hazardous Driving Scenario Detection for Supporting Autonomous Vehicles in Edge-Cloud Collaborative Environments

Ensuring the resilience of deep learning algorithms against adversarial attacks during edge-cloud data transmission between edge and cloud systems is crucial. Although significant strides have been made in enhancing the accuracy of hazardous driving scenario detection in autonomous vehicles, bolstering the robustness against adversarial attacks during data transfer and model recognition remains an urgent challenge. In this paper, we propose a novel approach to enhance adversarial robustness and maintain high accuracy in benign data. Using a lightweight CNN model, we detect hazardous driving scenarios, while a generative adversarial network generates decision boundary samples from original scenario images. These samples are incorporated into adversarial training, preventing overfitting on adversarial examples. Experimental results show our approach achieves robust accuracy (AUC) of 0.83 and 0.66 under FGSM and PGD attacks, surpassing vanilla adversarial training and TRADES methods with improving the standard accuracy (AUC) on benign samples by 0.22 and 0.12 relatively. Our proposed method advances beyond current adversarial training techniques to significantly enhance the model’s resilience to adversarial attacks during edge-cloud data transmission phases and minimize the loss of standard accuracy simultaneously. This advancement is crucial in enhancing the capability to correctly detect a wider range of hazardous driving scenarios, thereby providing significant support for the secure deployment of autonomous vehicles in edge-cloud collaborative environments.