Modeling the Abnormality: Machine Learning-based Anomaly and Intrusion Detection in Software-defined Networks
Loading...
Authors
Das, Tapadhir
Issue Date
2023
Type
Dissertation
Language
Keywords
Adversarial Attacks , Cybersecurity , Machine Learning , Network Intrusion Detection , Network Security , Software-defined Networks
Alternative Title
Abstract
Modern software-defined networks (SDN) provide additional control and optimal functionality over large-scale computer networks. Due to the rise in networking applications, cyber attacks have also increased progressively. Modern cyber attacks wreak havoc on large-scale SDNs, many of which are part of critical national infrastructures. Artifacts of these attacks may present as network anomalies within the core network or edge anomalies in the SDN edge. As protection, intrusion and anomaly detection must be implemented in both the edge and core. In this dissertation, we investigate and create novel network intrusion and anomaly detection techniques that can handle the next generation of network attacks. We collect and use new network metrics and statistics to perform network intrusion detection. We demonstrated that machine learning models like Random Forest classifiers effectively use network port statistics to differentiate between normal and attack traffic with up to 98% accuracy. These collected metrics are augmented to create a new open-sourced dataset that improves upon class imbalance. The developed dataset outperforms other contemporary datasets with an Fμ score of 94% and a minimum F score of 86%. We also propose SDN intrusion detection approaches that provide high confidence scores and explainability to provide additional insights and be implemented in a real-time environment. Through this, we observed that network byte and packet transmissions and their robust statistics can be significant indicators for the prevalence of any attack. Additionally, we propose an anomaly detection technique for time-series SDN edge devices. We observe precision and recall scores inversely correlate as ε increases, and ε = 6.0 yielded the best F score. Results also highlight that the best performance was achieved from data that had been moderately smoothed (0.8 ≤ α ≤ 0.4), compared to intensely smoothed or non-smoothed data. In addition, we investigated and analyzed the impact that adversarial attacks can have on machine learning-based network intrusion detection systems for SDN. Results show that the proposed attacks provide substantial deterioration of classifier performance in single SDNs, and some classifiers deteriorate up to ≈60. Finally, we proposed an adversarial attack detection framework for multi-controller SDN setups that uses inherent network architecture features to make decisions. Results indicate efficient detection performance achieved by the framework in determining and localizing the presence of adversarial attacks. However, the performance begins to deteriorate when more than 30% of the SDN controllers have become compromised. The work performed in this dissertation has provided multiple contributions to the network security research community like providing equitable open-sourced SDN datasets, promoting the usage of core network statistics for intrusion detection, proposing robust anomaly detection techniques for time-series data, and analyzing how adversarial attacks can compromise the machine learning algorithms that protect our SDNs. The results of this dissertation can catalyze future developments in network security.
Description
Citation
Publisher
License
Creative Commons Attribution 4.0 United States