Trust-Aware and Adaptive Malicious Node Detection in Fog Network using Independent DQN with Centralized Training

Abstract

The proliferation of decentralized and dynamic networks, such as fog computing and the Internet of Things (IoT), has significantly raised the demand for network security and resilience solutions. This study presents a decentralized trust management framework for detecting malicious nodes in fog-to-fog networks using Multi-Agent Reinforcement Learning (MARL). Our approach utilizes Independent Deep Q-Networks (I-DQN) for decentralized decision making at each fog node with dynamic trust evaluation, allowing them to learn an optimal detection policy based on local observations. Importantly, we enhance this by centralized training controlled by a central orchestrator, which uses a shared global critic and parameter sharing. The proposed system was evaluated on a 10-15 nodes fog network under three distinct attack scenarios: aggressive, stealth, and gradual. Experimental results demonstrate superior performance with detection rates of 92.0% for aggressive attacks, 78.0% for stealth attacks, and 67.9% for gradual attacks. Security focused results demonstrate exceptional false negative performance with FNR values of 8.3% ± 2.0% for aggressive attacks (excellent performance), 22.0% ± 2.0% for stealth attacks (good performance), and 31.9% ± 1.9% for gradual attacks (acceptable performance), ensuring minimal malicious nodes remain undetected across all attack types.   The proposed approach provides a highly secure and scalable solution for detecting malicious nodes in fog networks, offering superior threat detection through intelligent trust based decision making and coordinated multi-agent learning.