Improving Privacy in Federated Learning-Based Intrusion Detection for IoT Networks

Federated learning has emerged as a leading machine learning paradigm, promising to preserve data privacy while collaboratively training models. This approach is increasingly finding applications in the Internet of Things, particularly in the context of intrusion detection. However, most federated learning-based intrusion detection systems focus on detecting attacks that target or originate from network devices. Regrettably, federated learning systems are not immune to attacks by adversaries aimed at inferring information about device data. During model updates, learning processes and aggregation, malicious aggregators have the potential to infer information about client data. Secure aggregation is used to collect and consolidate model updates from various devices, ensuring privacy in individual contributions. However, current protocols face challenges such as high number of communication rounds, communication overhead and computation costs, all of which negatively impact model performance. This work aims to find the best balance between cost, privacy and effectiveness by studying different secure aggregation methods (multi-party computation, differential privacy and homomorphic encryption) to reach the best approach to improve privacy in intrusion detection systems based on federated learning for the Internet of Things, without affecting performance.