Federated Intrusion Detection on Non-IID Data for IIoT Networks Using Generative Adversarial Networks and Reinforcement Learning

Federated learning (FL) has become the promising approach for building collaborative intrusion detection systems (IDS) as providing privacy guaranteeing among data holders. Nevertheless, the non-independent and identically distributed (Non-HD) data in real-world scenarios negatively impacts the performance of aggregated models from training client updates. To this end, in this paper, we introduce Generative Adversarial Networks (GANs) and Reinforcement Learning (RL) approach for federated IDS that can deal with Non-HD data among organizational networks. More specifically, the imbalanced state between data classes is tackled by GAN-based data augmentation, while RL provides better performance in the client choosing process for federated IDS model training. Finally, the experimental results on Kitsune dataset indicate that our work can help to set up the collaboration between data holders for building more effective IDS to deploy in practice with distinguished data distribution.

Automated summary

This paper introduces a federated IDS approach using Generative Adversarial Networks (GANs) and Reinforcement Learning (RL) to handle non-independent and identically distributed data in organizational networks. The imbalanced data classes are addressed through GAN-based data augmentation, while RL improves the client choosing process for federated IDS model training. Experimental results on the Kitsune dataset demonstrate that this approach facilitates collaboration between data holders to build more effective IDS systems with distinguished data distribution.