跳转至

SegPAR：面向语义分割的以类别为中心的基于决策的稀疏攻击

文章背景与核心概要

语义分割作为计算机视觉中的一项核心任务，其安全性近年来受到了广泛关注。然而，尽管基于决策的稀疏对抗攻击在图像分类领域已得到充分研究，但在语义分割任务中却长期未被充分探索。现有的分类适配方法往往由于采用“以图像为中心”的像素累积策略而导致严重的查询效率低下，在广阔的图像空间中极大地浪费了查询预算。

为了填补这一空白，本文提出了 SegPAR 框架。该框架创新性地引入了“以类别为中心的探索范式”，将优化焦点从全图范围转移到特定类别的目标上；同时，设计了一种全新的差异奖励机制（Discrepancy Reward），消除了像素累积过程中由标准决策奖励产生的误导性反馈。大量实验表明，SegPAR 在稀疏效率和平均交并比（MIoU）降低方面显著优于现有的黑盒基线方法，其性能甚至可与白盒稀疏攻击相媲美。

📋 摘要

SegPAR is a novel decision-based black-box adversarial attack framework designed for semantic segmentation tasks. While sparse decision-based threats are well-studied in image classification, they have been largely underexplored in semantic segmentation.

SegPAR 是一个专为语义分割任务设计的创新的基于决策的黑盒对抗攻击框架。虽然基于稀疏决策的威胁在图像分类中已经得到了充分的研究，但在语义分割中却在很大程度上未被充分探索。

Existing classification-adapted methods often suffer from severe query inefficiency due to image-centric pixel accumulation, which wastes query budgets across vast image spaces. SegPAR resolves this by introducing: 1. A Class-Centric Exploration Paradigm: Shifting focus from image-wide optimization to class-specific targets. 2. A Novel Discrepancy Reward: Eliminating misleading feedback typically generated by standard decision rewards during pixel accumulation.

现有的分类适配方法往往由于以图像为中心的像素累积而遭受严重的查询效率低下，这在广阔的图像空间中浪费了查询预算。SegPAR 通过引入以下内容解决了这个问题： 1. 以类别为中心的探索范式： 将重点从全图优化转向特定类别的目标。 2. 新颖的差异奖励： 消除了在像素累积过程中由标准决策奖励通常产生的误导性反馈。

Experimental results demonstrate that SegPAR significantly outperforms existing black-box baselines in sparsity efficiency and Mean Intersection over Union (MIoU) reduction, achieving performance competitive with white-box sparse attacks.

实验结果表明，SegPAR 在稀疏效率和平均交并比（MIoU）降低方面明显优于现有的黑盒基线，实现了与白盒稀疏攻击相竞争的性能。

📌 元数据 (Metadata)

  • arXiv ID: arXiv:2608.11285 [cs.CV]
  • Authors: Dongsu Song, DaeYun Go, Boseung Seo, Jay Hoon Jung
  • Primary Subject: Computer Vision and Pattern Recognition (cs.CV)
  • Other Subjects: Artificial Intelligence (cs.AI), Cryptography and Security (cs.CR), Machine Learning (cs.LG)
  • Conference/Venue: ECCV 2026 poster
  • Submission Date: August 11, 2026

  • Official Code Repository: GitHub - KAU-QuantumAILab/SegPAR

  • arXiv ID: arXiv:2608.11285 [cs.CV]

  • 作者: Dongsu Song, DaeYun Go, Boseung Seo, Jay Hoon Jung
  • 主要学科: 计算机视觉与模式识别 (cs.CV)
  • 其他学科: 人工智能 (cs.AI)、密码学与安全 (cs.CR)、机器学习 (cs.LG)
  • 会议/场地: ECCV 2026 墙报展示 (poster)
  • 提交日期: 2026年8月11日
  • 官方代码库: GitHub - KAU-QuantumAILab/SegPAR

📄 摘要原文 (Abstract)

Despite the practical relevance of sparse decision-based black-box threats, they have received limited attention in semantic segmentation. To bridge this gap, we adapt the most representative decision-based black-box sparse attacks from the classification domain to serve as baselines, establishing a rigorous benchmark for this underexplored setting. In this context, we demonstrate that one of the existing methods suffers from severe query inefficiency due to its image-centric pixel accumulation, which rapidly exhausts query budgets across the vast image space. To overcome this, we propose SegPAR, a novel decision-based framework that shifts to a class-centric exploration paradigm. Furthermore, to eliminate the misleading feedback generated by standard decision rewards during pixel accumulation, we introduce a novel discrepancy reward. Extensive experiments show that SegPAR significantly outperforms black-box baselines in sparsity efficiency and MIoU reduction, while remaining competitive with white-box sparse attacks.

尽管稀疏的基于决策的黑盒威胁具有很高的实际相关性，但它们在语义分割中受到的关注却很有限。为了弥补这一空白，我们将分类领域中最具代表性的基于决策的黑盒稀疏攻击改造为基线，为这一未充分探索的设置建立了一个严密的基准。在此背景下，我们证明了现有的一种方法由于其以图像为中心的像素累积而遭受严重的查询效率低下，这在广阔的图像空间中迅速耗尽了查询预算。为了克服这一问题，我们提出了 SegPAR，这是一个新颖的基于决策的框架，它转变为以类别为中心的探索范式。此外，为了消除在像素累积过程中由标准决策奖励产生的误导性反馈，我们引入了一种新颖的差异奖励。大量实验表明，SegPAR 在稀疏效率和 MIoU 降低方面显著优于黑盒基线，同时在白盒稀疏攻击中保持了竞争力。

🔗 获取与资源 (Access & Resources)