Energy Optimization in IoT Adaptive Security: a Performance Comparison of Deep Reinforcement Learning Approaches

The Internet of Things (IoT) is revolutionizing industries, yet security remains a major concern. One that is tied to energy efficiency, especially in distributed and mobile IoT environments where resource constraints amplify the challenge. Adaptive security has emerged as a promising solution, dynamically adjusting protection levels in response to evolving threats. However, these adaptations often come at a high energy cost that accelerates device depletion and shortens network lifespan. In our previous work, we introduced a Deep Reinforcement Learning (DRL)-based mobility-aware framework using Deep QNetworks (DQN) with an SDN(Software-Defined Networking) based agent. It optimizes security-energy trade-offs for adaptive security solutions. In this paper, we extend the evaluation and algorithmic scope of the approach. While the original DQNbased solution established the foundational framework, this work advances the research in two critical dimensions. First, we introduce three new variants of the algorithm using Proximal Policy Optimization (PPO), Trust Region Policy Optimization (TRPO), and Advantage Actor-Critic (A2C) methods. Second, we conduct a comprehensive comparative analysis of all four approaches (DQN, PPO, TRPO, and A2C) under identical experimental conditions. Our results demonstrate that the PPO variant achieves superior performance. It outperforms both the original DQN implementation and the other policy-gradient methods in energy efficiency and network lifetime, while maintaining stable convergence in dynamic threat environments.