VIFL: vulnerability identification using federated learning in the internet of things systems

Vulnerability identification has been broadly studied as a way to improve cybersecurity. Internet of Things (IoT) ecosystems are considered particularly vulnerable as a whole, due to their widespread deployment, low processing ability, difficulty in updating and patching, and links to the physical world. Federated Learning (FL)-based vulnerability identification is still emerging in maturity and has limitations, including learning from heterogeneous data, detecting vulnerabilities in IoT networks, and model parameter privacy leakage. To address these limitations, we propose a novel FL-based vulnerability identification approach, named VIFL, which utilises variance reduction and staleness factor embedded with differential privacy (DP) in a semi-synchronous manner. VIFL is evaluated using the IoT Kitsune, BoT-IoT, and ToN-IoT datasets. The results show that the proposed approach achieved a 90.82% accuracy on Kitsune dataset, which is higher than DP-FedAvg, FedProx, DP-SCAFFOLD, and FedNova by an average of 30%. Furthermore, VIFL achieved effective performance while training with differential privacy (DP), even including the presence of stragglers. Thus, VIFL presents a promising approach for identifying vulnerabilities in IoT ecosystems and can be used to enhance their security. The source code for VIFL can be found on GitHub.