FedSBS: Federated-Learning participant-selection method for Intrusion Detection Systems

• Published in: Computer networks (Amsterdam, Netherlands : 1999) Netherlands : 1999), 2024-05, Vol.244, p.110351, Article 110351
• Main Authors: Cunha Neto, Helio N., Hribar, Jernej, Dusparic, Ivana, Fernandes, Natalia C., Mattos, Diogo M.F.
• Format: Article
• Language: English
• Subjects: Collaborative learning; Federated Learning; Intrusion Detection System; Machine learning
• ISSN: 1389-1286; 1872-7069
• DOI: 10.1016/j.comnet.2024.110351

Federated Learning (FL) is a decentralized machine learning approach in which multiple participants collaboratively train a model. Participants keep data locally, train their local models, and aggregate them in a single global model in a federated server. Collaborative FL-based Intrusion Detection Systems face challenges on an uneven statistical distribution of data and malicious participants trying to subvert the learning process. The statistical hurdles associated with imbalanced data and malicious participants pose a risk of skewing the training with biased or random data. The inability to effectively manage these statistical inconsistencies may degrade system performance, leading to false intrusion detection or opening avenues for cybersecurity breaches. To overcome these challenges, we propose a training method that employs score-based participant selection and utilizes global momentum for model aggregation. Our method improves the global model performance while mitigating the risks posed by malicious participants. The proposal incorporates a scoring system based on an information gain variant to evaluate each participant’s contribution. The scoring system and an epsilon greedy selection method ensure robust participant selection in each aggregation round. Furthermore, incorporating a global momentum term helps preserve previous knowledge at each aggregation round, contributing to model stability and overall learning. The proposed solution has demonstrated superior performance, delivering 80% F1-Score and 90% accuracy on experiments even in the presence of malicious participants, revealing the robustness and effectiveness of the proposal in mitigating statistical challenges. Consequently, the proposed method significantly enhances the performance of federated learning models, leading to more secure and efficient collaborative intrusion detection systems.