INCREMENTAL LEARNING FOR EDGE NETWORK INTRUSION DETECTION
Alina Florina Glavan, Victor Croitoru · REVUE ROUMAINE DES SCIENCES TECHNIQUES — SÉRIE ÉLECTROTECHNIQUE ET ÉNERGÉTIQUE · 2023
The paper presents incremental learning as a solution for adapting intrusion detection systems to the dynamic edge network conditions. Extreme gradient boost trees are proposed and evaluated with the Network Security Laboratory - Knowledge Discovery in Databases (NSL-KDD) benchmark dataset. The accuracy of the XGBoost classifier model improves by 15% with 1% of the KDD-test+ data used for training. A mechanism based on unsupervised learning that triggers retraining of the XGBoost classifier is suggested. These results are relevant in the context of model retraining on resource scarce environments (relative to a cloud environment), such as the network edge or edge devices.