Scholarlink

<p>Deep neural networks (DNNs) are vulnerable to backdoor attacks, where the model produces incorrect outputs when presented with inputs containing specific triggers. However, existing trigger detection methods often suffer from high implementation complexity and significant computational overhead. This paper proposes a novel and lightweight approach by embedding customized triggers in the frequency domain to effectively detect DNN backdoors. By embedding a custom frequency-domain trigger into potentially untrustworthy images and analyzing the model’s responses, it becomes possible to differentiate between clean and poisoned inputs. Experimental results demonstrate that our method achieves average detection rates of 97.47%, 99.40%, and 99.31% on the Fashion-MNIST, CIFAR-10, and EuroSAT-RGB datasets, respectively. This efficient detection strategy provides a practical and scalable solution to enhance DNN security against backdoor attacks.</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

Embedding customized triggers in frequency domain for simplified and efficient detection of deep neural network backdoors

  • 期刊论文

  • 发表日期：

    2026年7月23日

查看全文
  • Zigang Chen,
  • Zhangqi Wang,
  • Deyang Xie,
  • Yingrong Liu,
  • Haihua Zhu,
  • Dongfeng Yuan

摘要

Deep neural networks (DNNs) are vulnerable to backdoor attacks, where the model produces incorrect outputs when presented with inputs containing specific triggers. However, existing trigger detection methods often suffer from high implementation complexity and significant computational overhead. This paper proposes a novel and lightweight approach by embedding customized triggers in the frequency domain to effectively detect DNN backdoors. By embedding a custom frequency-domain trigger into potentially untrustworthy images and analyzing the model’s responses, it becomes possible to differentiate between clean and poisoned inputs. Experimental results demonstrate that our method achieves average detection rates of 97.47%, 99.40%, and 99.31% on the Fashion-MNIST, CIFAR-10, and EuroSAT-RGB datasets, respectively. This efficient detection strategy provides a practical and scalable solution to enhance DNN security against backdoor attacks.