Scholarlink

<p>Cyber-physical systems (CPS), particularly Robot Operating System (ROS) based robotic platforms are widely deployed in safety critical engineering environments, where reliable monitoring and intrusion detection are essential. The existing intrusion detection approaches for cyber-physical systems remain limited by poor generalization, high false positive rates, and inability to effectively model complex temporal–spatial attack behaviors. This highlight a critical need for an intelligent and adaptive framework that can enhance detection accuracy while maintaining robustness for real-time CPS environments. This paper proposes a compact hybrid deep learning framework that combines a convolutional neural network (CNN) front end for per window feature extraction with temporal encoders including bidirectional Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM) and a lightweight Transformer, supported by a streaming oriented preprocessing pipeline. The preprocessing stage applies median imputation for numerical features, top-K encoding for high cardinality categorical attributes and incremental dimensionality reduction to enable real-time processing. Model robustness is enhanced through curriculum based data augmentation and adversarial training using Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD). The proposed CNN + GRU and CNN + Transformer models are evaluated on the full ROSPaCe dataset using class preserving splits and standard multi class intrusion detection metrics. Both architectures achieve approximately 96% detection accuracy, with the CNN + GRU model consistently delivering higher precision, recall and F1-score across diverse attack classes. A deployment oriented analysis shows that while the Transformer better captures long range temporal dependencies at a modest computational cost, the GRU based model provides a superior accuracy efficiency trade off, making it more suitable for real time ROS deployments. Overall, the proposed CNN + GRU framework offers a robust, efficient, and practically deployable solution for intrusion detection in ROS based cyber physical systems.</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

A Hybrid Deep Learning Framework for Robust Monitoring and Intrusion Detection in Cyber-Physical Systems

  • 期刊论文

  • 发表日期：

    2026年6月3日

查看全文
  • Adeela Hayat,
  • Tahir Alyas,
  • Qaiser Abbas,
  • Adnan Nadeem,
  • Hani Almoamari,
  • Nasser Aljohani,
  • Nadia Tabassum

摘要

Cyber-physical systems (CPS), particularly Robot Operating System (ROS) based robotic platforms are widely deployed in safety critical engineering environments, where reliable monitoring and intrusion detection are essential. The existing intrusion detection approaches for cyber-physical systems remain limited by poor generalization, high false positive rates, and inability to effectively model complex temporal–spatial attack behaviors. This highlight a critical need for an intelligent and adaptive framework that can enhance detection accuracy while maintaining robustness for real-time CPS environments. This paper proposes a compact hybrid deep learning framework that combines a convolutional neural network (CNN) front end for per window feature extraction with temporal encoders including bidirectional Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM) and a lightweight Transformer, supported by a streaming oriented preprocessing pipeline. The preprocessing stage applies median imputation for numerical features, top-K encoding for high cardinality categorical attributes and incremental dimensionality reduction to enable real-time processing. Model robustness is enhanced through curriculum based data augmentation and adversarial training using Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD). The proposed CNN + GRU and CNN + Transformer models are evaluated on the full ROSPaCe dataset using class preserving splits and standard multi class intrusion detection metrics. Both architectures achieve approximately 96% detection accuracy, with the CNN + GRU model consistently delivering higher precision, recall and F1-score across diverse attack classes. A deployment oriented analysis shows that while the Transformer better captures long range temporal dependencies at a modest computational cost, the GRU based model provides a superior accuracy efficiency trade off, making it more suitable for real time ROS deployments. Overall, the proposed CNN + GRU framework offers a robust, efficient, and practically deployable solution for intrusion detection in ROS based cyber physical systems.