Personalized Few-Shot Federated Meta-Learning with Transfer Knowledge for Zero-Day Attack Detection in Resource-Constrained Wireless Sensor Security Under 6G THz Networks

Abstract

Wireless sensor networks of IoT have very relevant security threats regarding zero-day attacks, new attacks, and no training patterns, which put conventional detection to the test. Not only does this pose a challenge to detection since there are few labeled samples once the zero-day attacks are detected (5–20), but also limited power and processing resources, in addition to privacy matters in decentralized settings. We present a state-of-the-art solution based on personalized federated meta-learning and few-shot learning. Our solutions combine federated learning (FL) for privacy-preserving decentralized training, model-agnostic meta-learning (MAML) for few-shot learning adaptation, and transfer learning (TL) for prior exposure to the attacks. We implement a lightweight model (12.79 KB) with a personalized layer, meaning that while the model is trained globally during federated training, each sensor node can also adapt to its specific local network features. We validate our solution on CICIDS2017, which includes four completely unknown zero-day attack types: Bot, DoS Slowloris, Heartbleed, and DoS GoldenEye. We achieve 64.04% accuracy and 77.93% F1-Score in the 20-shot scenario, 467% greater than the baseline (11.29% accuracy) while achieving 100% precision and size of the model (25–66 times smaller than the rest). Our results prove that the combination of FL, MAML, and TL is an effective solution for few-shot detection of zero-day attacks in real IoT networks, where conventional solutions cannot operate with such extreme limitations.