A Novel Neural Networks-based Framework for APT Detection in Networked Autonomous Systems

Hassan El Alami, Danda B. Rawat

Abstract: Networked Autonomous Systems (AS) represent a networking paradigm where nodes autonomously make decisions, operating independently or collaboratively to achieve network goals. Unlike traditional networks, networked AS emphasizes decentralized control, enabling nodes to adapt dynamically without relying heavily on centralized controllers. This autonomy promotes agility, scalability, and resilience, making networked AS applicable to various domains like IoT and wireless communications. However, this architecture is vulnerable to various threats, including Advanced Persistent Threat (APT), which exploits vulnerabilities leading to resource depletion and hindering legitimate network operations. APTs, particularly prevalent in IoT-enabled networked AS, pose a significant risk by overwhelming network resources and disrupting services. To address these challenges, this paper introduces a new neural network-based framework tailored to detecting APT in networked AS environments. The proposed framework leverages generative adversarial networks and Deep Learning (DL) to accurately identify malicious activities in networked AS. We evaluated our framework using the newly released dataset TON_IoT, which contains a comprehensive variety of APT and addresses gaps in existing datasets in the context of AS. We obtained a significant improvement in APT detection by comparing DL techniques such as Conventional Neural Networks, Gated Recurrent Units, Long Short-Term Memory, Multilayer Perceptron (MLP), and combining MLP with Auto-Encoder (MLP+AE). The proposed framework demonstrates important accuracy and robustness by analyzing the selected DL techniques’ performance as influenced by variations in the training set size.