FedRand: A Federated Random Forest Learning Technique for Anomaly Detection in IoT Networks

Federated Learning (FL) is an emerging distributed machine learning (ML) technique distinguished by non-independent and identically distributed (Non-IID) data, statistical heterogeneity, and an expected large number of participating clients to collaboratively train a shared model without fusing the training data into one centralized server. State-of-the-art FL research focuses on gradient-based models, which is not suitable for ML-based intrusion detection systems that utilize tree-based learning methods such as Random Forest. Adapting a typical gradient-based FL method to a tree-based training technique is non-trivial, as ensembling trees and aggregating decision trees from different random forests across clients can be computationally, spatially, and temporally intensive. To overcome these challenges, this paper proposes FedRand, a novel adaptive Federated Random Forest Aggregation Learning Technique for Anomaly Detection in Internet of Things (IoT) networks. This paper thoroughly examines a suite of novel tree selection and aggregation strategies within a federated learning framework, ensuring robust model accuracy, accelerated aggregation, and global model convergence. We believe that this work opens up a promising solution for federated tree-based learning techniques.