Continual Learning for Anomaly based Network Intrusion Detection

Amalapuram, Suresh Kumar and Tadwai, Akash and Vinta, Reethu and Channappayya, Sumohana and Tamma, Bheemarjuna Reddy (2022) Continual Learning for Anomaly based Network Intrusion Detection. In: 14th International Conference on COMmunication Systems and NETworkS, COMSNETS 2022, 4 January 2022 through 8 January 2022, Bangalore.

[img] Text
2022_14th_International_Conference_on_COMmunication_Systems_and_NETworkS2.pdf - Published Version
Restricted to Registered users only

Download (1MB) | Request a copy


To defend computing systems against ever-growing cyber attacks, Anomaly-based Network Intrusion Detection Systems (A-NIDS) have to evolve continuously. This requirement renders classical machine learning algorithms ineffective since they do not handle sequentially evolving tasks gracefully. Specifically, neural networks (NNs) are prone to catastrophic forgetting (CF) when trained on sequential data. Recent advances in addressing this drawback of NNs have resulted in a paradigm called Continual Learning (CL) which mitigates CF by introducing suitable constraints during the sequential training of these NNs. CL has been shown to be very effective in improving the performance of NNs on computer vision tasks. However, its application to the design of A-NIDS has not been explored. In this work, we evaluate the suitability of CL to address the challenges posed in A-NIDS design. Unlike computer vision datasets, network datasets suffer from the Class Imbalance (CI) problem, which makes the direct application of CL algorithms challenging. To evaluate the suitability of CL algorithms on network datasets, we study the impact of class imbalance on task ordering and its effect on the design of CL- based A-NIDS in the Class Incremental (CIL) and Domain Incremental (DIL) learning settings. Towards this end, we apply two popular CL algorithms viz. Elastic Weight Consolidation (EWC) and Gradient Episodic Memory (GEM) on two datasets viz., CICIDS and KDD Cup'99, and evaluate their performance. We found that CI affects task order sensitivity to a greater extent in the CIL setting when compared to the DIL setting. The performance of DIL setting can be further enhanced by incorporating experience forgetting aware memory population techniques, and we recommend this as a practical approach to building CL-based A-NIDS. © 2022 IEEE.

[error in script]
IITH Creators:
IITH CreatorsORCiD
Channappayya, Sumohana
Tamma, Bheemarjuna Reddy
Item Type: Conference or Workshop Item (Paper)
Uncontrolled Keywords: Anomaly based network intrusion detection systems (A-NIDS); Continual learning; Machine learning; Network datasets; Task order sensitivity
Subjects: Computer science
Electrical Engineering
Divisions: Department of Electrical Engineering
Depositing User: . LibTrainee 2021
Date Deposited: 25 Jul 2022 06:28
Last Modified: 25 Jul 2022 06:28
Publisher URL:
Related URLs:

Actions (login required)

View Item View Item
Statistics for RAIITH ePrint 9901 Statistics for this ePrint Item