Towards Improving the Anti-attack Capability of the RangeNet++

With the possibility of deceiving deep learning models by appropriately modifying images verified, lots of researches on adversarial attacks and adversarial defenses have been carried out in academia. However, there is few research on adversarial attacks and adversarial defenses of point cloud semantic segmentation models, especially in the field of autonomous driving. The stability and robustness of point cloud semantic segmentation models are our primary concerns in this paper. Aiming at the point cloud segmentation model RangeNet++ in the field of autonomous driving, we propose novel approaches to improve the security and anti-attack capability of the RangeNet++ model. One is to calculate the local geometry that can reflect the surface shape of the point cloud based on the range image. The other is to obtain a general adversarial sample related only to the image itself and closer to the real world based on the range image, then add it into the training set for training. The experimental results show that the proposed approaches can effectively improve the RangeNet + +’s defense ability against adversarial attacks, and meanwhile enhance the RangeNet++ model’s robustness.