BadCLIP: Dual-Embedding Guided Backdoor Attack on Multimodal Contrastive Learning

📄 arXiv: 2311.12075v3 📥 PDF

作者: Siyuan Liang, Mingli Zhu, Aishan Liu, Baoyuan Wu, Xiaochun Cao, Ee-Chien Chang

分类: cs.CV

发布日期: 2023-11-20 (更新: 2024-03-04)

备注: The paper lacks some work that needs to be cited

期刊: CVPR 2024


💡 一句话要点

提出BadCLIP以解决多模态对比学习中的后门攻击问题

🎯 匹配领域: 支柱二:RL算法与架构 (RL & Architecture) 支柱九:具身大模型 (Embodied Foundation Models)

关键词: 后门攻击 多模态对比学习 安全性 对抗性防御 视觉触发模式

📋 核心要点

  1. 现有的后门攻击方法在多模态对比学习模型中容易被专门的防御机制抵消,存在有效性不足的问题。
  2. 本文提出了一种双嵌入引导的后门攻击框架,确保视觉触发模式与文本目标语义在嵌入空间中接近,从而提高攻击的隐蔽性。
  3. 实验结果显示,本文方法在面对最先进的后门防御时,攻击成功率提升了45.3%,有效性显著高于现有方法。

📝 摘要(中文)

研究后门攻击对模型版权保护和增强防御具有重要价值。尽管现有后门攻击已成功感染多模态对比学习模型(如CLIP),但它们容易被针对MCL模型的专门防御所抵消。本文揭示了在实际场景中,后门攻击在防御后仍然有效的威胁,并介绍了 oolns攻击,该攻击对后门检测和模型微调防御具有抵抗力。我们从贝叶斯规则的角度出发,提出了一种双嵌入引导的后门攻击框架,确保视觉触发模式在嵌入空间中接近文本目标语义,从而使得后门学习引起的细微参数变化难以被检测。大量实验表明,我们的攻击在面对最先进的后门防御时,显著超越了现有基线(+45.3% ASR),几乎使这些缓解和检测策略失效。

🔬 方法详解

问题定义:本文旨在解决多模态对比学习模型中的后门攻击问题,现有方法容易被专门的防御机制识别和抵消,导致攻击效果不佳。

核心思路:我们提出了一种双嵌入引导的后门攻击框架,通过确保视觉触发模式与文本目标语义在嵌入空间中接近,来提高后门攻击的隐蔽性和有效性。

技术框架:该框架主要包括两个模块:一是视觉触发模式的生成,二是对生成模式的优化,使其与目标视觉特征对齐,从而增强攻击效果。

关键创新:本文的核心创新在于引入双嵌入引导机制,使得后门攻击在防御机制下仍然有效,且难以被检测,与现有方法相比具有更高的隐蔽性和攻击成功率。

关键设计:在设计中,我们采用了特定的损失函数来优化视觉触发模式,并通过调整参数设置来确保触发模式与目标语义的紧密结合,从而增强攻击的有效性。

🖼️ 关键图片

fig_0
fig_1
fig_2

📊 实验亮点

实验结果表明,本文提出的BadCLIP攻击在面对最先进的后门防御时,攻击成功率提升了45.3%,显著超越了现有的基线方法,几乎使得这些防御策略失效,展示了其在实际应用中的威胁性。

🎯 应用场景

该研究的潜在应用领域包括安全性要求高的多模态学习系统,如图像与文本结合的内容生成、自动驾驶系统中的多模态感知等。通过提高后门攻击的隐蔽性,研究结果将推动对抗性防御机制的发展,增强模型的安全性和鲁棒性。

📄 摘要(原文)

Studying backdoor attacks is valuable for model copyright protection and enhancing defenses. While existing backdoor attacks have successfully infected multimodal contrastive learning models such as CLIP, they can be easily countered by specialized backdoor defenses for MCL models. This paper reveals the threats in this practical scenario that backdoor attacks can remain effective even after defenses and introduces the \emph{\toolns} attack, which is resistant to backdoor detection and model fine-tuning defenses. To achieve this, we draw motivations from the perspective of the Bayesian rule and propose a dual-embedding guided framework for backdoor attacks. Specifically, we ensure that visual trigger patterns approximate the textual target semantics in the embedding space, making it challenging to detect the subtle parameter variations induced by backdoor learning on such natural trigger patterns. Additionally, we optimize the visual trigger patterns to align the poisoned samples with target vision features in order to hinder the backdoor unlearning through clean fine-tuning. Extensive experiments demonstrate that our attack significantly outperforms state-of-the-art baselines (+45.3% ASR) in the presence of SoTA backdoor defenses, rendering these mitigation and detection strategies virtually ineffective. Furthermore, our approach effectively attacks some more rigorous scenarios like downstream tasks. We believe that this paper raises awareness regarding the potential threats associated with the practical application of multimodal contrastive learning and encourages the development of more robust defense mechanisms.