A Federated Learning Framework for Collaborative and Privacy-Preserving IoT Security Attack Prediction

Abstract

The rapid expansion of the Internet of Things (IoT) has multiplied the attack surface exposed to cyber threats. Traditional intrusion detection systems rely on collecting network data at a central location, which raises serious privacy, bandwidth, and regulatory concerns in IoT environments where devices are heterogeneous, resource-constrained, and geographically distributed. Federated Learning (FL) offers a promising alternative by enabling devices to collaboratively train a shared model without sharing their raw data. This thesis investigates how a federated learning-based framework can be designed to enable distributed IoT devices to collaboratively predict security attacks, while preserving data privacy and handling device heterogeneity. We propose FedShield, a federated learning framework for multi-class IoT attack prediction, built around a lightweight deep learning classifier, Federated Averaging, and a mask-based secure aggregation mechanism. The framework is implemented in Python using TensorFlow and the Flower library, and evaluated on the TON_IoT dataset partitioned across non-IID clients. The experimental results show that the framework reaches strong detection performance on a non-IID partition of the TON_IoT dataset, and that mask-based secure aggregation can be added without measurable utility cost. The work confirms that federated learning is a credible building block for privacy-preserving IoT intrusion detection.