Scholarlink

<p>Illumination variations pose a pervasive challenge for vision models deployed in real-world scenarios. Previous works on illumination robustness have primarily focused on naturalness rather than physical interpretability, exacerbating the disparity between the physical and digital worlds. This paper proposes an adversarial relighting framework, IL-Attack, which maliciously manipulates incident light through inverse rendering (IR) and physically based rendering (PBR). The intrinsic decoupling ability of IR and the physical interpretability of PBR facilitate the realism of our attack, offering a potential solution for physical attacks. Inspired by natural effects like occlusions, we introduce an optimizable soft mask that can be applied to the incidence, compatible with any ray-tracing technique and capable of mimicking natural light effects such as neon lighting and shadow casting. Extensive experiments validate that our methods generate natural adversarial relighting results with physical interpretability. Simulated rendering results underscore the potential threats of adversarial relighting examples. Codes and datasets are available at <a href="https://github.com/zhongcz332/IL-Attack">https://github.com/zhongcz332/IL-Attack</a>.</p>

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

Adversarial relighting attacks: physically interpretable manipulation of incident light for vision model vulnerability exploration

  • 期刊论文

  • 发表日期：

    2025年7月21日

查看全文
  • Chengzhi Zhong,
  • Mengda Xie,
  • Yiling He,
  • Ping Li,
  • Meie Fang

摘要

Illumination variations pose a pervasive challenge for vision models deployed in real-world scenarios. Previous works on illumination robustness have primarily focused on naturalness rather than physical interpretability, exacerbating the disparity between the physical and digital worlds. This paper proposes an adversarial relighting framework, IL-Attack, which maliciously manipulates incident light through inverse rendering (IR) and physically based rendering (PBR). The intrinsic decoupling ability of IR and the physical interpretability of PBR facilitate the realism of our attack, offering a potential solution for physical attacks. Inspired by natural effects like occlusions, we introduce an optimizable soft mask that can be applied to the incidence, compatible with any ray-tracing technique and capable of mimicking natural light effects such as neon lighting and shadow casting. Extensive experiments validate that our methods generate natural adversarial relighting results with physical interpretability. Simulated rendering results underscore the potential threats of adversarial relighting examples. Codes and datasets are available at https://github.com/zhongcz332/IL-Attack.