Scholarlink

Deep neural networks have begun to receive large amounts of point cloud data for tasks such as intelligent driving and robot recognition. Meanwhile, neural networks have been proven to be highly susceptible to adversarial perturbations. Most existing whitebox-based perturbation attack methods require attackers to fully understand the architecture and parameters of the victim model, and often stay away from the decision boundary of the model for high attack success rate. This complex requirement makes it difficult to achieve in real application scenarios. In order to solve this problem, we will focus on the research of point cloud disturbance transferability. Studying the transferability of perturbation attacks is of great significance for enhancing the security and reliability of deep neural networks in real-world application scenarios. We propose a new method for point cloud perturbation attack called DBL-Attack. Based on the decomposition perturbation method, the model prediction is treated as a probability distribution, and a decision boundary loss term is introduced to optimize the original loss function. This ensures that adversarial perturbations are kept away from the decision boundary during the update process, further improving their transferability. Our method consistently demonstrates superior performance across both real-world and synthetic datasets.

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

DBL-Attack: Decision Boundary Loss-Driven Adversarial Attack to Generate Transferable Traffic Object Point Clouds Using Factorized Random Perturbation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Min Xie,
  • Weiquan Liu,
  • Xingwang Huang,
  • Jinhe Su,
  • Zongyue Wang,
  • Guorong Cai

摘要

Deep neural networks have begun to receive large amounts of point cloud data for tasks such as intelligent driving and robot recognition. Meanwhile, neural networks have been proven to be highly susceptible to adversarial perturbations. Most existing whitebox-based perturbation attack methods require attackers to fully understand the architecture and parameters of the victim model, and often stay away from the decision boundary of the model for high attack success rate. This complex requirement makes it difficult to achieve in real application scenarios. In order to solve this problem, we will focus on the research of point cloud disturbance transferability. Studying the transferability of perturbation attacks is of great significance for enhancing the security and reliability of deep neural networks in real-world application scenarios. We propose a new method for point cloud perturbation attack called DBL-Attack. Based on the decomposition perturbation method, the model prediction is treated as a probability distribution, and a decision boundary loss term is introduced to optimize the original loss function. This ensures that adversarial perturbations are kept away from the decision boundary during the update process, further improving their transferability. Our method consistently demonstrates superior performance across both real-world and synthetic datasets.