Scholarlink

With the growing prominence of point cloud recognition, ensuring robustness against natural degradations and adversarial perturbations remains a critical challenge. Existing methods often fall short in effectively addressing these issues. In this work, we propose PointCMI, a novel adversarial training framework that leverages conditional mutual information (CMI) to achieve robust point cloud recognition. Specifically, we introduce a two-step strategy: (1) extracting class-specific feature centers from a clean model, and (2) constraining the perturbed outputs to remain close to these centers during adversarial training. By using CMI as a principled metric, PointCMI explicitly captures the class-wise complexity of clean and adversarial point clouds. This leads to a novel CMI minimization objective that drives both robustness and accuracy. Extensive experiments on ModelNet40 and ShapeNetPart demonstrate that PointCMI significantly outperforms existing defenses, setting a new benchmark in robust point cloud recognition.

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

Information-Theoretic Point Cloud Defense: Harnessing Conditional Mutual Information Against Adversarial Attacks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Kaiwen Zhang,
  • Xinhao Zhong,
  • Shuoyang Sun,
  • Jiaxin Hong,
  • Bin Chen,
  • Teko Ranoka,
  • Xuan Wang,
  • Shutao Xia

摘要

With the growing prominence of point cloud recognition, ensuring robustness against natural degradations and adversarial perturbations remains a critical challenge. Existing methods often fall short in effectively addressing these issues. In this work, we propose PointCMI, a novel adversarial training framework that leverages conditional mutual information (CMI) to achieve robust point cloud recognition. Specifically, we introduce a two-step strategy: (1) extracting class-specific feature centers from a clean model, and (2) constraining the perturbed outputs to remain close to these centers during adversarial training. By using CMI as a principled metric, PointCMI explicitly captures the class-wise complexity of clean and adversarial point clouds. This leads to a novel CMI minimization objective that drives both robustness and accuracy. Extensive experiments on ModelNet40 and ShapeNetPart demonstrate that PointCMI significantly outperforms existing defenses, setting a new benchmark in robust point cloud recognition.