A Novel Approach for Intrusion Detection using Online Federated Learning on Streaming Data

This paper studies the application of online federated learning for intrusion detection systems. An experiment is conducted with two different models, Gaussian Naive Bayes and Semi-supervised Federated Learning on Evolving Data Streams (SFLEDS), which are evaluated in four different settings, centralized offline, centralized online, federated offline, and federated online. The models are evaluated on the NSL-KDD dataset, and the federated models are run with 20,30, and 40 clients. The results show that for Naive Bayes, the centralized offline models have the best performance, while for SFLEDS, the federated online models perform the best with accuracy scores around 90%. Suggestions for improvements of the models are provided in the discussion, with the conclusion being that, while the results show promising results for federated online learning when employed for intrusion detection systems, the models used need to be carefully selected to achieve good results. Further research is also required for different models, such as deep learning models, which might achieve even better results.