A Dynamic Spatiotemporal Deep Learning Solution for Cloud–Edge Collaborative Industrial Control System Distributed Denial of Service Attack Detection

Zhigang Cao, Bo Liu, Dayong Gao, Ding Zhou, Xiaopeng Han, Jiuxin CaoElectronics 2025 · DOI

TL;DR

This paper introduces FedDynST, a novel cloud-edge collaborative system for detecting Distributed Denial of Service (DDoS) attacks in Industrial Control Systems (ICS). By combining federated learning with dynamic and static spatiotemporal graph neural networks, FedDynST effectively captures complex traffic patterns while preserving data privacy, achieving superior detection accuracy and faster convergence than existing methods.

What Problem Does This Paper Address?

Industrial Control Systems (ICS) are increasingly integrated with IT systems and cloud-edge architectures, creating new vulnerabilities. DDoS attacks pose a significant threat to these systems, potentially causing production halts, service interruptions, and even safety risks. Existing DDoS detection methods often struggle to effectively handle the unique characteristics of ICS traffic and the distributed nature of cloud-edge environments. Furthermore, centralizing sensitive ICS data for training detection models raises privacy concerns. This paper aims to address these challenges by developing a privacy-preserving, accurate, and efficient DDoS detection solution tailored for cloud-edge collaborative ICS.

Methods

The core of the proposed solution is the FedDynST model, which integrates two key components:

  • Federated Learning Framework: This allows multiple ICS clients to collaboratively train a global DDoS detection model without sharing their raw traffic data. Each client trains a local model, and only model updates are sent to a central server for aggregation. To improve the learning process, the paper introduces a dynamic weighting algorithm that assigns weights to client updates based on the divergence between local and global model parameters and the client's local training loss. This ensures that higher-quality client models contribute more to the global model.
  • Dynamic Spatiotemporal Deep Learning Model: This model uses a graph convolutional network (GCN) approach, specifically the APPNP (Attention-based Propagation Network) variant, to analyze traffic features.
  • Static and Dynamic Adjacency Matrices: To capture both long-term correlations and short-term anomalies in ICS traffic, the model constructs two types of adjacency matrices. The static matrix represents enduring relationships, while the dynamic matrix captures real-time changes, crucial for detecting sudden attacks.
  • Graph Convolution: The APPNP layer uses these matrices to learn complex relationships between traffic features.
  • 1D-CNN Layer: This layer extracts temporal sequence features from the traffic data.
  • Pooling and Fully Connected Layers: These layers further process the extracted features for final classification.
  • Unsupervised Loss Function: An unsupervised loss function based on maximizing mutual information is used to enhance the correlation between outputs from static and dynamic graph convolutions, improving feature representation.

Datasets

The FedDynST model was validated using two publicly available datasets:

  • CICDDoS2019: A widely used benchmark dataset for DDoS detection, containing traffic data from 13 common DDoS attack types and normal traffic.
  • Edge-IIoTset: A more recent benchmark dataset designed for IoT and IIoT security research, which includes industrial traffic data using the Modbus/TCP protocol, making it highly relevant for ICS security studies.

The datasets were preprocessed by aggregating bidirectional flows into sub-flow samples and then divided into training, validation, and test sets for each client. A global test set was also created to evaluate the overall model generalization.

Key Insights and Ideas

  • Dynamic Spatiotemporal Analysis is Crucial for ICS DDoS Detection: The paper highlights that simply looking at static traffic patterns or individual temporal features isn't enough. Capturing the interplay between long-term ICS behavior (static matrix) and sudden attack-induced anomalies (dynamic matrix) is key to effective detection.
  • Federated Learning is Essential for Privacy in ICS: The distributed nature of ICS and the sensitivity of their data make federated learning a natural fit. It allows for collaborative model improvement without compromising data privacy.
  • Dynamic Weighting Improves Federated Learning: Standard federated averaging can be suboptimal when client data quality varies. The proposed dynamic weighting algorithm, which considers model divergence and local loss, leads to faster convergence and better global model performance.
  • Mutual Information Maximization Enhances Feature Representation: Using mutual information to maximize the correlation between features learned from static and dynamic graph convolutions helps the model better understand the underlying relationships in the traffic data.

Key Results

The FedDynST model demonstrated superior performance across both datasets and various evaluation metrics (accuracy, precision, recall, F1-score, AUC).

  • CICDDoS2019 Dataset:
    • Achieved an average accuracy of 0.9932 on the global test set, outperforming other models.
  • Client detection accuracy consistently exceeded 0.96, showing strong adaptability.
  • Demonstrated a 2.57% improvement in accuracy over the FLAD+CNN model.
  • Edge-IIoTset Dataset:
    • Achieved an accuracy of 0.9947 on the global test set.
  • Showed excellent performance on this dataset, which is more representative of real-world ICS environments.
  • Convergence: The FedDynST model exhibited faster convergence speed compared to other federated learning models, reaching high accuracy in fewer rounds.
  • Parameter Sensitivity: The model's performance was sensitive to the weight of the dynamic adjacency matrix (η), with an optimal value around 0.3, and the weight of the unsupervised loss (μ), with an optimal value around 0.1.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that while the FedDynST model shows strong performance, the effectiveness of the dynamic weighting algorithm is influenced by the quality of local client data.
  • The study focused on DDoS attacks; extending the model to detect other types of cyber threats in ICS could be beneficial.
  • Future Work:
    • The authors suggest investigating the engineering implementation of DDoS attack detection in cloud-edge collaborative ICSs for practical deployment.
  • Further research could focus on enhancing identification capabilities across different industrial contexts and developing more targeted defense strategies.

Summary updated

Abstract

With the continuous development of industrial intelligence, the integration of cyber–physical components creates a need for effective attack detection methods to mitigate potential DDoS threats. Although several DDoS attack detection modeling approaches have been proposed, few effectively incorporate the unique characteristics of industrial control system (ICS) architectures and traffic patterns. This paper focuses on DDoS attack detection within cloud–edge collaborative ICSs and proposes a novel detection model called FedDynST. This model combines federated learning and deep learning to construct feature graphs of traffic data. Introducing dynamic and static adjacency matrices, this work reveals the interactions between long-term industrial traffic data and short-term anomalies associated with DDoS attacks. Convolutional neural networks are utilized to capture distinctive temporal features within industrial traffic, thereby improving the detection precision. Moreover, the model enables continuous optimization of the global detection framework through a federated learning-based distributed training and aggregation mechanism, ensuring the privacy and security of industrial client data. The effectiveness of the FedDynST model was validated on the CICDDoS2019 and Edge-IIoTset datasets. The simulation results validated the superiority of the proposed approach, and thus, demonstrated significant improvements in both detection accuracy and convergence.
BibTeX
@inproceedings{cao2025dynamic,
  author = {Zhigang Cao and Bo Liu and Dayong Gao and Ding Zhou and Xiaopeng Han and Jiuxin Cao},
  title = {{A Dynamic Spatiotemporal Deep Learning Solution for Cloud–Edge Collaborative Industrial Control System Distributed Denial of Service Attack Detection}},
  booktitle = {Electronics},
  year = {2025},
  doi = {10.3390/electronics14091843},
}
Figure