Fusion‑Based Deep Reinforcement Learning for High‑Precision Network Intrusion Classification

Abstract

The increasing sophistication of cyber-attacks demands intelligent intrusion detection systems capable of adapting to dynamic and evolving threat landscapes. This study presents an advanced intrusion detection framework using the Network Intrusion dataset, integrating deep reinforcement learning and ensemble-based fusion techniques. Data preprocessing includes handling missing values, encoding categorical attributes, and standardizing numerical features to ensure consistency and learning stability. Multiple reinforcement learning algorithms are explored, including Deep Q Network, Deep Deterministic Policy Gradient, Proximal Policy Optimization, Twin Delayed Deep Deterministic Policy Gradient, and Soft Actor-Critic. To further enhance detection capability, a Fusion Model combining DQN, TD3, and Naive Bayes is proposed, along with an Extension Fusion Model integrating DQN, TD3, and a Voting Classifier. Experimental evaluation demonstrates that the Extension Fusion model achieves superior performance with 0.9988 accuracy, 0.9993 precision, 0.9465 recall, and a 0.9705 F1-score, outperforming individual models. For real-world deployment, a Flask-based web application is developed with secure user interaction, real-time feature input, automated preprocessing, and prediction visualization. The system classifies network traffic into Normal, DoS, R2L, Probe, and U2R categories, providing an accurate, scalable, and interpretable intrusion detection solution