Scholarlink

<p>Cyber-physical systems (CPS) are a key part of modern infrastructure, blending physical and computational control almost seamlessly. Robust security, especially via intrusion detection systems (IDS), iscritical. In our study, we present an IDS for CPS that combines federated learning with the SimpleNet deep learning model. Federated learning boosts data privacy and security by training 10 local models then merging them into one global model without the need toexchange raw data among nodes. We selected SimpleNet because of its simplicity and its ability to handle complex datasets; our data prep involved removing duplicates, addressing missing values, and doing feature selection with a Random Forest classifier. We evaluated our approach on the TON_IoT and NSL-KDD datasets by trying different feature subsets. The top performing variant achieved an accuracy over 99% for ToN-IoT and over 98% for NSL-KDD. Compared to existing solutions, our IDS stands out with its high accuracy, efficient feature selection, and robustness in a federated learning context, making it well-suited for edge computing environments. Furthermore, this method supports data privacy and scalability, effectively tackling the key challenges current IDS approaches face.</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

Enhancing data privacy in cyber-physical systems with federated learning-based intrusion detection

  • 期刊论文

  • 发表日期：

    2026年6月15日

查看全文
  • Kamal Bella,
  • Azidine Guezzaz,
  • Vinayakumar Ravi,
  • Said Benkirane,
  • Mouaad Mohy-eddine,
  • Mourade Azrour,
  • Slimane Ennajar

摘要

Cyber-physical systems (CPS) are a key part of modern infrastructure, blending physical and computational control almost seamlessly. Robust security, especially via intrusion detection systems (IDS), iscritical. In our study, we present an IDS for CPS that combines federated learning with the SimpleNet deep learning model. Federated learning boosts data privacy and security by training 10 local models then merging them into one global model without the need toexchange raw data among nodes. We selected SimpleNet because of its simplicity and its ability to handle complex datasets; our data prep involved removing duplicates, addressing missing values, and doing feature selection with a Random Forest classifier. We evaluated our approach on the TON_IoT and NSL-KDD datasets by trying different feature subsets. The top performing variant achieved an accuracy over 99% for ToN-IoT and over 98% for NSL-KDD. Compared to existing solutions, our IDS stands out with its high accuracy, efficient feature selection, and robustness in a federated learning context, making it well-suited for edge computing environments. Furthermore, this method supports data privacy and scalability, effectively tackling the key challenges current IDS approaches face.