A Federated Learning-POMDP Approach for Network Intrusion Detection

Due to the increasing expanse of the internet, cyber attacks have become more prevalent, thus necessitating the need for robust security controls, especially with respect to machine learning-assisted intrusion detection. A recent method for intrusion detection involves federated learning, which ensures the confidentiality of data as opposed to traditional, centralized machine learning. In this study, we propose a novel framework involving the formulation of a federated learning network intrusion detection system as a partially observable Markov decision process (POMDP). We then solve this POMDP by applying a reinforcement learning algorithm, specifically the asynchronous advantage actor critic algorithm. Our results indicate that our A3C statistically performs better compared with our benchmark federated learning models and obtains a higher recall score while maintaining a high precision score. Henceforth, our proposed model decreases the number of false negatives while maintaining high overall performance. Our novel, unique framework is a promising approach for federated network intrusion detection.