A study on Few-shot Learning approach for Intrusion Detection System with Class Incremental Learning
Qui Phan Xuan Cao, Duong Dai Tran, Son Tran Thai Ngo, Nghi Hoang Khoa, Phan The Duy, Van-Hau Pham · 2025
Traditional machine learning-based intrusion detection systems are difficult to adapt to new attacks.Over time, these models become less effective due to outdated knowledge, making it challenging to detect emerging cyberattacks.On the other hand, the resourceconsuming retraining process raises concerns that the model will forget about the old attacks, which are more likely to happen in cyberspace.Meta-learning methods offer promising solutions by adapting to new security challenges, such as Zero-day attacks, despite limited samples, without concerns about catastrophic interference.We combine meta-learning with incremental learning to present a Few-shot Class Incremental Learning method for Deep Neural Networks.This method enables the model to learn about new attacks quickly while retaining the stability to recognize previously learned types, thereby enhancing its ability to detect various threats.The experimental results on the CIC-IDS2017 dataset show that this model is efficient in detecting both old and new intrusion techniques.