首页|基于引导扩散模型的自然对抗补丁生成方法

基于引导扩散模型的自然对抗补丁生成方法

扫码查看
近年来,物理世界中的对抗补丁攻击因其对深度学习模型安全的影响而引起了广泛关注.现有的工作主要集中在生成在物理世界中攻击性能良好的对抗补丁,没有考虑到对抗补丁图案与自然图像的差别,因此生成的对抗补丁往往不自然且容易被观察者发现.为了解决这个问题,本文提出了一种基于引导的扩散模型的自然对抗补丁生成方法.具体而言,本文通过解析目标检测器的输出构建预测对抗补丁攻击成功率的预测器,利用该预测器的梯度作为条件引导预训练的扩散模型的逆扩散过程,从而生成自然度更高且保持高攻击成功率的对抗补丁.本文在数字世界和物理世界中进行了广泛的实验,评估了对抗补丁针对各种目标检测模型的攻击效果以及对抗补丁的自然度.实验结果表明,通过将所构建的攻击成功率预测器与扩散模型相结合,本文的方法能够生成比现有方案更自然的对抗补丁,同时保持攻击性能.
A Guided Diffusion-based Approach to Natural Adversarial Patch Gen-eration
Adversarial patch attacks in the physical world have gained a lot of attention in recent years due to their safety implications.Existing work has mostly focused on generating adversarial patches that can attack certain models in the physical world,but the resulting patterns are often unnatural and easy to identify.To tackle this problem,we propose a guided diffusion-based approach to natural adversarial patch generation.Specifically,we construct a predictor for attack success rate(ASR)prediction by parsing the output of the target detector,such that the reverse process of a pre-trained diffu-sion model can be guided by the gradient of the classifier to generate adversarial patches with improved naturalness and high ASR.We conduct extensive experiments in both the digital and the physical worlds to evaluate the attack effective-ness against various object detection models,as well as the naturalness of generated patches.The experimental results show that by combining the ASR predictor with a pre-trained diffusion model,our method is able to produce more natural adver-sarial patches than the state-of-art approaches while remaining highly effective.

object detectionadversarial patchdiffusion modeladversarial exampleadversarial attackdeep learning

何琨、佘计思、张子君、陈晶、汪欣欣、杜瑞颖

展开 >

武汉大学国家网络安全学院,湖北武汉 430072

武汉大学空天信息安全与可信计算教育部重点实验室,湖北武汉 430072

武汉大学日照信息技术研究院,山东日照 276800

地球空间信息技术协同创新中心,湖北武汉 430079

展开 >

目标检测 对抗补丁 扩散模型 对抗样本 对抗攻击 深度学习

国家重点研发计划中央高校基本科研业务费专项国家自然科学基金国家自然科学基金湖北省重点研发计划山东省重点研发计划

2022YFB31021002042022kf103462206203620761872022BAA0392022CXPT055

2024

电子学报
中国电子学会

电子学报

CSTPCD北大核心
影响因子:1.237
ISSN:0372-2112
年,卷(期):2024.52(2)
  • 42