| Texto completo | |
| Autor(es): |
Sameera, K. M.
;
Vinod, P.
;
Rocha, Anderson
;
Rehiman, K. A. Rafidha
;
Conti, Mauro
Número total de Autores: 5
|
| Tipo de documento: | Artigo Científico |
| Fonte: | JOURNAL OF INFORMATION SECURITY AND APPLICATIONS; v. 95, p. 19-pg., 2025-10-28. |
| Resumo | |
The rapid growth of Internet of Things (IoT) devices has expanded the cyber-attack surface, making traditional Network Intrusion Detection Systems (NIDS) less effective against modern, dynamic threats. The rise of privacy concerns and legal restrictions also limits the use of centralized security systems, highlighting the need for decentralized alternatives. Federated Learning (FL)-based NIDS addresses this by training models without sharing private user data. However, these systems are still vulnerable to poisoning attacks and can suffer from performance issues due to varied client data. In this paper, we introduce WeiDetect, a novel two-phase defense mechanism for FL-based NIDS. Operating on the server side, WeiDetect tackles both adversarial attacks and client data heterogeneity. It works by evaluating local models with a validation dataset, fitting their performance scores to a Weibull distribution for identifying and excluding malicious or low-quality models before aggregation. Our experimental results show that WeiDetect outperforms existing defenses, improving target class recall by up to 70% and enhancing the global model's F1 score by 1%-14%. (AU) | |
| Processo FAPESP: | 23/12865-8 - Horus: técnicas de inteligência artificial para detecção e análise de realidades sintéticas |
| Beneficiário: | Anderson de Rezende Rocha |
| Modalidade de apoio: | Auxílio à Pesquisa - Temático |