Deep Reinforcement Learning for Intelligent Cybersecurity in Smart City IoT Infrastructures
Rajarshi Tarafdar, Harmeet Singh, Vijay Pahuja, Gaurav Garg, Ramaswamy Sivaraman, B. Jegajothi · 2025
Cybersecurity threats are evolving rapidly, necessitating intelligent and adaptive defense mechanisms. This paper proposes a Deep Reinforcement Learning (DRL)-based cybersecurity framework that leverages a Deep Q-Network (DQN) for real-time intrusion detection and mitigation. The model transforms raw network traffic into structured state representations, enabling effective learning of attack patterns over time. The action space is designed to implement various countermeasures, including traffic blocking, source isolation, and alert triggering. A reward function is formulated to balance detection accuracy while minimizing false positives and missed attacks. The proposed DQN-based security agent undergoes training with experience replay and target network stabilization to optimize its decision-making policies. Experimental evaluations on benchmark datasets demonstrate superior performance, achieving an accuracy of 98.4 %, outperforming conventional machine learning models such as LSTM, Random Forest, and SVM. Additionally, the model exhibits high detection rates for diverse cyber threats, including DoS, ransomware, botnets, phishing, and SQL injection. Comparative analysis highlights the effectiveness of reinforcement learning in adaptive threat mitigation, offering a scalable and intelligent security solution for modern network infrastructures.