Asynchronous Advantage Actor-Critic (A3C) Learning for Cognitive Network Security
Eric Muhati, Danda B. Rawat
Abstract
Undoubtedly, the recent implacable, widespread, and intricate cyber-attacks demand cognitive cyber-defense techniques. Although machine learning (ML) and artificial intel-ligence (AI) algorithms are extensively applied for enhanced network security, existing AI-based agents are impeded by incongruous shifts in actionable intelligence. For instance, the low-cost network data extraction prerequisites are crucial for AI-based network security research, yet seldom. Seemingly, there is increased urgency for unconstrained network data access to improve cyber-AI efficiency in unfamiliar threat scenarios. While host-based attacks can be detected from the analysis of extracted data, network-based attacks require adaptive network traffic ex-amination. This paper proposes an automated network scanning and data-mining technique through open-source service discovery tools for deep reinforcement learning (DRL) based cognitive network intrusion detection system (NIDS). Our proposed DRL-NIDS is developed using an asynchronous advantage actor-critic (A3C) AI method that demonstrates improved results compared to related works. We combine the best parts of predicting both the value and the optimal policy functions in extensive experimentation with 3 datasets, namely UNSW-NB15, AWID, and NSL-KDD. Our experiment shows degraded performance from other state-of-the-art DRL-based NIDS, while our proposed A3C-NIDS achieves a 98.68% accuracy with the least false alarm rate.