An Adversarial DBN-LSTM Method for Detecting and Defending against DDoS Attacks in SDN Environments

Lei Chen, Zhihao Wang, Ru Huo, Tao HuangAlgorithms 2023 · DOI

TL;DR

This paper introduces an adversarial Deep Belief Network-Long Short-Term Memory (DBN-LSTM) approach to detect and defend against Distributed Denial of Service (DDoS) attacks in Software-Defined Networking (SDN) environments. By leveraging Generative Adversarial Networks (GANs) to create adversarial training data, their method aims to improve the robustness of intrusion detection systems against sophisticated evasion techniques.

What Problem Does This Paper Address?

Cyber Ranges, critical for network security research and training, are highly susceptible to DDoS attacks. Traditional Machine Learning (ML)-based Network Intrusion Detection Systems (NIDSs) are increasingly vulnerable to adversarial attacks, where attackers subtly modify attack traffic to evade detection. This poses a significant challenge for accurately identifying and mitigating these threats, especially in dynamic SDN environments. The paper aims to address this by developing a more resilient NIDS that can effectively handle both standard and adversarially crafted DDoS attacks.

Methods

The proposed system is built around an adversarial DBN-LSTM framework, comprising four key modules:

  • Data Collection: The SDN controller gathers IP flow records from physical and virtual switches in the data layer.
  • Data Processing: Non-numerical features are converted to numerical ones, and original numerical features are normalized. Crucially, the Shannon entropy formula is used to evaluate the centralization of IP addresses and port numbers, a characteristic that can change during DDoS attacks.
  • Adversarial Deep Learning Anomaly Detection: This is the core of the proposed method. It utilizes a Generative Adversarial Network (GAN) to generate adversarial samples. These samples, along with normal and unperturbed DDoS samples, are used to train a DBN-LSTM classifier. The GAN's generator creates fake data to fool the discriminator, while the discriminator learns to distinguish real from fake. This adversarial training makes the DBN-LSTM model more robust to adversarial perturbations. The DBN component is used for dimensionality reduction and feature extraction, while the LSTM component captures temporal dependencies in the network traffic.
  • Abnormal Defending: Upon detecting a DDoS attack, this module implements an exception prevention mechanism, such as dropping malicious packets, based on event-condition-action rules.

Datasets

The experiments were conducted using the public CICDDoS2019 dataset. This dataset contains both normal traffic and various types of DDoS attacks. The authors note that the dataset was preprocessed to extract over 80 features, resulting in a large number of records. For adversarial testing, they generated a new testing dataset using the FGSM algorithm on the CICDDoS2019 data.

Key Insights and Ideas

  • Adversarial Training is Crucial for Robustness: The paper strongly emphasizes that standard ML/DL models are easily fooled by adversarial attacks. By incorporating GAN-generated adversarial samples into the training process, the DBN-LSTM model becomes significantly more resilient.
  • DBN for Feature Extraction and Dimensionality Reduction: The use of Deep Belief Networks (DBNs) before the LSTM layer is highlighted as a way to efficiently extract features and reduce the dimensionality of the data, thereby lowering computational costs and improving scalability.
  • LSTM for Temporal Dynamics: Long Short-Term Memory (LSTM) networks are essential for capturing the time-series nature of network traffic, which is critical for identifying attack patterns that evolve over time.
  • Shannon Entropy for Traffic Centralization: The application of Shannon entropy to analyze the distribution of IP addresses and ports is an interesting approach to identify anomalies indicative of concentrated attack traffic.

Key Results

The proposed adversarial DBN-LSTM method achieved strong performance on the CICDDoS2019 dataset:

  • Accuracy: 96.55%
  • Precision: 96.44%
  • Recall: 98.53%
  • F1 Score: 97.47%

When tested against adversarial DDoS attacks generated using the FGSM algorithm, the method demonstrated a significant improvement in accuracy (91.23%) compared to DBN-LSTM and LSTM models without GAN, which performed poorly (less than 8%). The paper also found that a four-layer LSTM structure provided the best performance.

Limitations and Future Work

  • Data Availability: The authors state that the data is not publicly available due to privacy concerns, which could limit reproducibility for some aspects.
  • Computational Cost: While DBN is used for dimensionality reduction, the overall adversarial training process, especially with GANs, can still be computationally intensive.
  • Generalizability: The paper focuses on SDN environments and specific types of DDoS attacks. Future work could explore the method's effectiveness across a wider range of network architectures and attack vectors. The authors suggest further research into improving DL algorithms or combining them with other relevant technologies to enhance performance in SDN environments.

Summary updated

Abstract

As an essential piece of infrastructure supporting cyberspace security technology verification, network weapons and equipment testing, attack defense confrontation drills, and network risk assessment, Cyber Range is exceptionally vulnerable to distributed denial of service (DDoS) attacks from three malicious parties. Moreover, some attackers try to fool the classification/prediction mechanism by crafting the input data to create adversarial attacks, which is hard to defend for ML-based Network Intrusion Detection Systems (NIDSs). This paper proposes an adversarial DBN-LSTM method for detecting and defending against DDoS attacks in SDN environments, which applies generative adversarial networks (GAN) as well as deep belief networks and long short-term memory (DBN-LSTM) to make the system less sensitive to adversarial attacks and faster feature extraction. We conducted the experiments using the public dataset CICDDoS 2019. The experimental results demonstrated that our method efficiently detected up-to-date common types of DDoS attacks compared to other approaches.
BibTeX
@inproceedings{chen2023adversarial,
  author = {Lei Chen and Zhihao Wang and Ru Huo and Tao Huang},
  title = {{An Adversarial DBN-LSTM Method for Detecting and Defending against DDoS Attacks in SDN Environments}},
  booktitle = {Algorithms},
  year = {2023},
  doi = {10.3390/a16040197},
}
Figure