Adaptive Meta-Learning for Zero-Day Threat Discovery in Autonomous Cyber Defences

Abstract

The proliferation of cyber threats has made traditional, signature-based and rule-based defence systems to be insufficient against zero-day attacks which exploit unknown vulnerabilities. In this paper, we present an Adaptive Meta-Learning Framework for zero day threat detection in hyper intelligent ACDSA. By incorporating the core elements of meta-learning, ADMoRe can learn how to learn; allowing rapid model adaptation across dynamic threat scenarios with a limited amount of labeled data. The designed system combines reinforcement learning for optimal decision making, graph neural networks for pattern recognition of related threat and federated learning to share intelligence in a decentralized manner within the network. Contrary to traditional novel anomaly detectors, the method is enable to generalize across domains where models trained on a class can also rapidly spot new unseen patterns. We demonstrate on popular benchmark cybersecurity datasets that our adaptive meta-learner achieves up to 94% in detection accuracy for zero-day exploits and reduces false alarms by 27%, when compared with state-of-the-art deep learning methods. The research highlights the disruptive potential of adaptive meta-learning for developing self-adaptive, autonomous defence ecosystems which can spontaneously mitigate zero-day threats in real-time even in highly dynamic networked cyber environments.