Evaluating and Improving the Local Robustness of Deep Neural Networks in the Domain of Aerial Perception

Semantic segmentation models are highly vulnerable to adversarial attacks, highlighting the need for improved Local Robustness. Local robustness reflects how robust the prediction of a trained model is with respect to input perturbations. To address the lack of local robustness in the context of Deep Learning, we apply a robust training scheme to the Unified Perceptual Parsing for Scene Understanding (UPerNet) architecture. In addition, we extend the Cross-Lipschitz Extreme Value for nEtwork Robustness (CLEVER) estimation to the segmentation task. The dataset of interest is the AeroScapes dataset, containing images and the corresponding semantic segmentation mask from the perspective of a drone. The used training scheme significantly improves the local robustness against white box adversarial attacks while maintaining the baseline performance of the model. The CLEVER evaluation further substantiates these improvements and suggests future directions for better qualitative evaluation of local robustness and the robust training scheme.