Abstract:

Existing intrusion detection techniques often require numerous malicious samples for model training.However, in real-world scenarios, only a small number of intrusion traffic samples can be obtained, which belong to few-shot scenarios.To address this challenge, a network intrusion detection method designed for few-shot scenarios was proposed.The method comprised two main parts: a packet sampling module and a meta-learning module.The packet sampling module was used for filtering, segmenting, and recombining raw network data, while the meta-learning module was used for feature extraction and result classification.Experimental results based on three few-shot datasets constructed from real network traffic data sources show that the method exhibits good applicability and fast convergence and effectively reduces the occurrence of outliers.In the case of 10 training samples, the maximum achievable detection rate is 99.29%, while the accuracy rate can reach a maximum of 97.93%.These findings demonstrate a noticeable improvement of 0.12% and 0.37% respectively, in comparison to existing algorithms.

Key words: intrusion detection, few-shot, meta-learning, network security, deep learning