Publication

Abstract : The increasing prevalence of cyber threats in wireless sensor networks (WSNs) necessitates adaptive and efficient security mechanisms to ensure robust data transmission while addressing resource constraints. This paper proposes a reinforcement learning-based adaptive encryption framework that dynamically scales encryption levels based on real-time network conditions and threat classification. The proposed model leverages a deep learning-based anomaly detection system to classify network states into low, moderate, or high threat levels, which guides encryption policy selection. The framework integrates dynamic Q-learning for optimizing energy efficiency in low-threat conditions and double Q-learning for robust security adaptation in high-threat environments. A Hybrid Policy Derivation Algorithm is introduced to balance encryption complexity and computational overhead by dynamically switching between these learning models. The proposed system is formulated as a Markov Decision Process (MDP), where encryption level selection is driven by a reward function that optimizes the trade-off between energy efficiency and security robustness. The adaptive learning strategy employs an ϵ-greedy exploration-exploitation mechanism with an exponential decay rate to enhance convergence in dynamic WSN environments. The model also incorporates a dynamic hyperparameter tuning mechanism that optimally adjusts learning rates and exploration parameters based on real-time network feedback. Experimental evaluations conducted in a simulated WSN environment demonstrate the effectiveness of the proposed framework, achieving a 30.5% reduction in energy consumption, a 92.5% packet delivery ratio (PDR), and a 94% mitigation efficiency against multiple cyberattack scenarios, including DDoS, black-hole, and data injection attacks. Additionally, the framework reduces latency by 37% compared to conventional encryption techniques, ensuring minimal communication delays. These results highlight the scalability and adaptability of reinforcement learning-driven adaptive encryption in resource-constrained networks, paving the way for real-world deployment in next-generation IoT and WSN applications.