Scholarlink

In recent years, adversarial attacks and adversarial training methods have become increasingly popular ways to improve the robustness of neural network algorithms. By generating adversarial examples through minor perturbations to the original samples, it is possible to obtain visually indistinguishable data on which the neural network predicts an erroneous result. This approach has been shown to enhance the quality of image classification in both supervised and unsupervised scenarios. The concept of using adversarial attack augmentation to improve the quality of a detector was confirmed on the VisDrone2019 dataset. In this paper, we introduce a new data augmentation and training method that enables us to achieve higher quality in domain adaptation problems, using object detection tasks with the YOLOv5 algorithm and the Cityscapes dataset, with Foggy Cityscapes data as the target domain. We utilized common approaches to generate adversarial examples and then trained the model on both clear and perturbed data. Additionally, we developed our adversarial attack method as a new data augmentation approach, referred to as AdvGrads. The results demonstrate that the proposed method exhibits superior quality compared to baseline approach.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Domain Adaptation for Object Detection via Adversarial Attack Augmentation

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Andrey Nikitin,
  • Vadim Gorbachev,
  • Stepan Syrovatkin

摘要

In recent years, adversarial attacks and adversarial training methods have become increasingly popular ways to improve the robustness of neural network algorithms. By generating adversarial examples through minor perturbations to the original samples, it is possible to obtain visually indistinguishable data on which the neural network predicts an erroneous result. This approach has been shown to enhance the quality of image classification in both supervised and unsupervised scenarios. The concept of using adversarial attack augmentation to improve the quality of a detector was confirmed on the VisDrone2019 dataset. In this paper, we introduce a new data augmentation and training method that enables us to achieve higher quality in domain adaptation problems, using object detection tasks with the YOLOv5 algorithm and the Cityscapes dataset, with Foggy Cityscapes data as the target domain. We utilized common approaches to generate adversarial examples and then trained the model on both clear and perturbed data. Additionally, we developed our adversarial attack method as a new data augmentation approach, referred to as AdvGrads. The results demonstrate that the proposed method exhibits superior quality compared to baseline approach.