TITLE:
Intelligent Approach to Intrusion Detection Based on Multi-Agent Reinforcement Learning (MARL) in Smart IoT Environments
AUTHORS:
Nguendap Severin, Fotsing Eric, Hamdane Allamine Moussa, Kakeu Tuekam Severin Vianey, Nzeukou Takougang Armand, Lienou Jean Pierre
KEYWORDS:
IoT Environments, Intrusion Detection, AE-LSTM, Reinforcement Learning, Multi-Agent, PPO, IoT-23
JOURNAL NAME:
International Journal of Intelligence Science,
Vol.15 No.4,
October
11,
2025
ABSTRACT: This work presents a new method to improve the security of Internet of Things (IoT) networks using a model combining an autoencoder LSTM (AE-LSTM) and a centralized multi-agent reinforcement learning (RL) system. IoT, essential in fields such as healthcare, smart cities, or agriculture, is facing increasing threats (DDoS attacks, intrusions, botnets like Mirai or Gafgyt) due to its often poorly protected connected objects. Traditional intrusion detection methods fail in the face of complex attacks, which has led to the exploration of artificial intelligence techniques. The AE-LSTM model extracts temporal features from network data, while the multi-agent reinforcement learning (RL) system, based on the PPO algorithm, enables collaborative and adaptive anomaly detection using common latent vectors. The IoT-23 dataset, realistic and captured in an IoT environment, was carefully prepared: data cleaning (removing missing and duplicate values), reducing high correlations, normalization, and class balancing. Experimental results show that the model achieves a precision of 99.30%, a recall of 99.35%, and an F1-score of 99.40%, with a reward curve associated with the PPO algorithm that shows constant progression during training, demonstrating stable convergence of the learning process. This approach outperforms other existing methods in terms of robustness and generalization and provides a solid foundation for real-time applications. The reward curve associated with the PPO algorithm shows constant progression during training, demonstrating stable convergence of the learning process. This approach outperforms other existing methods in terms of robustness and generalization and provides a solid foundation for real-time applications.