AI4FIDS: Multimodal Federated Intrusion Detection

The rapid progression of smart technologies creates several advantages like enhanced connectivity, personalisation solutions and environmental sustainability. However, this revolution creates also several cyber risks. In particular, the attackers have the ability to synthesise and automate advanced attack scenarios over time, while it is evident that Artificial Intelligence (AI) allows the composition of intelligent attack vectors that can adapt in real-time to conventional countermeasures. Despite the fact that AI can also benefit defensive mechanisms, there are still functional and privacy issues that need to be resolved. First, AI requires appropriate datasets that can differ from environment to environment. In addition, these datasets usually are not available due to privacy issues. Finally, adversarial attacks have the ability to target and affect the AI-based decision-making process. Therefore, in light of the previous remarks, we provide AI4FIDS, a multimodal Intrusion Detection System (IDS) for critical infrastructures. AI4FIDS leverages Federated learning (FL) and combines multiple data sources, thus allowing cooperative intelligence across multiple domains in a private manner and minimising the impact of potential adversarial attacks. In this paper, we present in detail the architectural design and specifications of AI4FIDS, while the evaluation results demonstrate their detection performance, taking into account several datasets and aggregation strategies. Finally, based on the evaluation results, we discuss how the overall reliability and detection capabilities (in terms of detecting multi-step attack scenarios) of AI4FIDS can be improved by combining the detection outcomes of the components behind AI4FIDS.