Scholarlink

With the rapid development of Artificial Intelligence & Internet of Things (AIoT), billions of intelligent devices are connected together, enabling smart control and convenient access. However, this interconnectedness also poses a serious risk of privacy breaches and malicious hijacking. As an emerging hardware security primitive, physical unclonable function (PUF) can be used for lightweight authentication. Yet, its vulnerability to modeling attacks necessitates complex design, leading to high costs and low reliability. To tackle this issue, this paper leverages the existing hardware in AIoT devices and proposes a neural network-based PUF (NN-PUF) protection method to obfuscate PUF responses, effectively enhancing PUF resilience against modeling attacks. Experimental results demonstrate that with a single-layer binary fully connected network of only 8 neurons, the PUF utilizing 16-bit responses remains secure even when the number of training data increases to 5.12 million, validating the effectiveness of our method against modeling attacks. Furthermore, due to the inherent fault tolerance of neural networks, our approach does not compromise reliability as much as traditional methods.

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

A Neural Network-Based PUF Protection Method Against Machine Learning Attack

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Haolin Jiang,
  • Chenxi Zhu,
  • Ding Deng,
  • Shengqiang Lou,
  • Pengyue Sun,
  • Lei Chen

摘要

With the rapid development of Artificial Intelligence & Internet of Things (AIoT), billions of intelligent devices are connected together, enabling smart control and convenient access. However, this interconnectedness also poses a serious risk of privacy breaches and malicious hijacking. As an emerging hardware security primitive, physical unclonable function (PUF) can be used for lightweight authentication. Yet, its vulnerability to modeling attacks necessitates complex design, leading to high costs and low reliability. To tackle this issue, this paper leverages the existing hardware in AIoT devices and proposes a neural network-based PUF (NN-PUF) protection method to obfuscate PUF responses, effectively enhancing PUF resilience against modeling attacks. Experimental results demonstrate that with a single-layer binary fully connected network of only 8 neurons, the PUF utilizing 16-bit responses remains secure even when the number of training data increases to 5.12 million, validating the effectiveness of our method against modeling attacks. Furthermore, due to the inherent fault tolerance of neural networks, our approach does not compromise reliability as much as traditional methods.