ProtoMAML: A Hybrid Meta-Learning Approach Integrating Prototypical Networks for Data-Efficient DDoS Attack Detection

Résumé fourni par la source

Distributed Denial of Service (DDoS) attacks remain a persistent and formidable threat, often overwhelming targets by saturating network bandwidth or exhausting server resources with massive volumes of malicious traffic. Traditional detection methods typically rely on signature-based approaches or large labeled training sets, posing challenges in rapidly changing attack landscapes where novel or variant DDoS vectors emerge frequently. To address this gap, we propose ProtoMAML, a hybrid meta-learning framework that integrates Prototypical Networks and Model-Agnostic Meta-Learning (MAML) to facilitate robust few-shot DDoS detection. By combining prototype-based clustering with fast, gradient-driven adaptation, ProtoMAML can accurately detect new attacks from only a handful of labeled flows per class. Experiments on a large-scale flow dataset (nearly half a million flows) demonstrate that ProtoMAML achieves a recall of up to 99.4% under severe data scarcity. Extended discussions on computational overhead, adversarial resilience, and real-world deployment provide insights into how meta-learning can offer a powerful, agile defense against evolving cyber threats.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.