2020/03/14 by Xiaogang Xu, Xu, Xiaogang, Hengshuang Zhao +3 · 4 citations
Computer Science · Medicine · #Adversarial Robustness in Machine Learning #Adversarial system #Algorithm #Anomaly Detection Techniques and Applications #Artificial intelligence #Artificial neural network #COVID-19 diagnosis using AI #Computer Vision and Pattern Recognition (cs.CV) #Computer science #Deep neural networks #Divide and conquer algorithms #FOS: Computer and information sciences #Inference #Machine learning #Pascal (unit) #Robustness (evolution) #Segmentation #Training (meteorology) #cs.CV
paper · pdf · doi:10.48550/arxiv.2003.06555
published in arXiv (Cornell University) (Cornell University) · ICCV2021
openalex publication_date 2020/03/14 · arxiv created 2021/08/16 · arxiv updated 2021/08/17 · openalex created_date 2025/10/10 · openalex updated_date 2026/08/05
Adversarial training is promising for improving robustness of deep neural networks towards adversarial perturbations, especially on the classification task. The effect of this type of training on semantic segmentation, contrarily, just commences. We make the initial attempt to explore the defense strategy on semantic segmentation by formulating a general adversarial training procedure that can perform decently on both adversarial and clean samples. We propose a dynamic divide-and-conquer adversarial training (DDC-AT) strategy to enhance the defense effect, by setting additional branches in the target model during training, and dealing with pixels with diverse properties towards adversarial perturbation. Our dynamical division mechanism divides pixels into multiple branches automatically. Note all these additional branches can be abandoned during inference and thus leave no extra parameter and computation cost. Extensive experiments with various segmentation models are conducted on PASCAL VOC 2012 and Cityscapes datasets, in which DDC-AT yields satisfying performance under both white- and black-box attack.