Utilizando aprendizado federado para detecção de intrusos em redes IOT

Abstract

The widespread adoption of the Internet of Things (IoT) has expanded the attack surface of computer networks, making security a central challenge in this ecosystem. Machine learning–based intrusion detection systems have shown promising results; however, centralized approaches require the transfer of sensitive data, raising concerns related to privacy and compliance with regulations such as the Brazilian General Data Protection Law (LGPD). In this context, Federated Learning emerges as an alternative by enabling collaborative model training without sharing raw data. This work evaluates the feasibility and performance of a Federated Learning–based Intrusion Detection System applied to IoT networks, considering realistic scenarios with severely imbalanced and non-independent and identically distributed (Non-IID) data. Experiments were conducted using the CICIoT2023 dataset and a deep neural network architecture optimized for federated environments, implemented with the Flower framework and the FedProx aggregation strategy. The results indicate that the federated model achieved a global accuracy of 82.80% and a macro F1-score of 81.20%, demonstrating competitive performance while preserving data privacy and ensuring feasibility for deployment on resource-constrained edge devices.