Class-Incremental Continual Learning for SDN-IoT Intrusion Detection: A Comparative Study of Replay Strategies under Limited Annotation Budgets
The integration of Software-Defined Networking (SDN) with the Internet of Things (IoT) creates new security challenges that require adaptive intrusion detection systems capable of learning emerging attack types without forgetting previously observed threats. This paper presents a systematic empirical study of class-incremental continual learning for SDN-IoT intrusion detection using the ASEADOS-SDN-IoT dataset. We compare three learning strategies, fine-tuning without replay, random experience replay, and class-balanced experience replay under three annotation budgets (5%, 10%, and 20% of new experience data). Experiments follow a realistic class-incremental schedule over Benign+DoS, DDoS, Bot, and Probe traffic. The results demonstrate that experience replay is essential to mitigate catastrophic forgetting: both random and class-balanced replay substantially outperform fine-tuning. At low-to-moderate budgets (5–10%), the two replay strategies achieve comparable macro-F1 scores (approximately 0.47–0.48). However, at a 20% budget, class-balanced replay suffers severe degradation, while random replay remains comparatively robust. Increasing the annotation budget beyond 10% yields diminishing or negative returns under a fixed memory constraint. The findings provide practical guidance for deploying memory-efficient, annotation-aware continual learning systems in operational SDN-IoT environments.
Authors
- J.Isabella
Publication Details
- Journal
- Dandao Xuebao/Journal of Ballistics
- Published
- 2026-11-16
- Primary Topic
- Network Security and Intrusion Detection
- Type
- article
- Field-Weighted Citation Impact
- 0.00