Scholarlink

<p>The growing reliance on deep neural networks (DNNs) has raised significant concerns about data security, particularly through backdoor attacks that compromise model security by embedding hidden triggers to mislead the model's predictions. State-of-the-art backdoor attacks have the following weaknesses: the visibility of their triggers, the need for compute-intense tasks to achieve invisible triggers, and the limited transferability across different models. In this paper, we propose PixVulNet, an invisible backdoor framework that exploits universal susceptible points within datasets, relying exclusively on the data itself and being unrelated to model details such as structure and gradients. Especially, our method is different from methods that use model attention mechanisms to find points. The susceptible points mentioned previously are called data-inherent vulnerabilities in our manuscript. The attacker only uses data distribution characteristics to create universal data-inherent vulnerabilities as triggers, which is invisible and requires less computation compared to state-of-the-art methods. In addition, we prove our method has a high cross-model transferability and is model agnostic, as it exploits training data itself. We also achieve a sparse adversarial attack by perturbing a few vulnerabilities (between 5 and 10 points per image) within a part of training samples (less than 10% of the data), with average attack success rates exceeding 50%. Therefore, our method enables the easy and efficient generation of triggers. Moreover, we demonstrate that data-inherent vulnerabilities cannot be removed by data augmentation. Extensive experiments on benchmark datasets and model structures verify the effectiveness of our method. The code will be available at <a href="https://github.com/esther11-max/PixVulNet/tree/master">https://github.com/esther11-max/PixVulNet/tree/master</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

Data-inherent vulnerabilities: an invisible framework for model-agnostic backdoor attacks

  • 期刊论文

  • 发表日期：

    2025年9月29日

查看全文
  • Mengru Liu,
  • Zhongguo Yang,
  • Yonglu Jiang,
  • Zhixin Wang,
  • Yuhan Ma,
  • Han Li

摘要

The growing reliance on deep neural networks (DNNs) has raised significant concerns about data security, particularly through backdoor attacks that compromise model security by embedding hidden triggers to mislead the model's predictions. State-of-the-art backdoor attacks have the following weaknesses: the visibility of their triggers, the need for compute-intense tasks to achieve invisible triggers, and the limited transferability across different models. In this paper, we propose PixVulNet, an invisible backdoor framework that exploits universal susceptible points within datasets, relying exclusively on the data itself and being unrelated to model details such as structure and gradients. Especially, our method is different from methods that use model attention mechanisms to find points. The susceptible points mentioned previously are called data-inherent vulnerabilities in our manuscript. The attacker only uses data distribution characteristics to create universal data-inherent vulnerabilities as triggers, which is invisible and requires less computation compared to state-of-the-art methods. In addition, we prove our method has a high cross-model transferability and is model agnostic, as it exploits training data itself. We also achieve a sparse adversarial attack by perturbing a few vulnerabilities (between 5 and 10 points per image) within a part of training samples (less than 10% of the data), with average attack success rates exceeding 50%. Therefore, our method enables the easy and efficient generation of triggers. Moreover, we demonstrate that data-inherent vulnerabilities cannot be removed by data augmentation. Extensive experiments on benchmark datasets and model structures verify the effectiveness of our method. The code will be available at https://github.com/esther11-max/PixVulNet/tree/master.