Search for a command to run...
Polymorphic adversarial agents that dynamically mutate their behavior present significant challenges to conventional intrusion detection systems, which often rely on static feature representations or fixed signature based models. Existing deep learning approaches, including CNNs, transformers, and graph neural networks (GNNs), demonstrate strong detection capabilities but exhibit limitations in handling continuously evolv ing attack patterns. In this paper, we propose an integrated entropy-driven agentic cyber defense model that combines graph neural network-based anomaly detection, generative adversarial network (GAN) poly morphic attack simulation, reinforcement learning-based mitigation, and game-theoretic attacker–defender modeling. It models the stochastic mutation of adversarial agents using entropy gradients, captures relational network structures through GNNs, simulates realistic polymorphic attacks using GANs, and dynamically adapts mitigation strategies via reinforcement learning. Experimental evaluation on the CICIoT2023 dataset demonstrates that the proposed model achieves superior performance over existing state-of-the-art methods, with an accuracy of 98.3%, F1-score of 0.98, and significant improvements in robustness against polymor phic attacks compared to CNN-based (93.2% accuracy, 0.92 F1), transformer-based (95.4%, 0.94 F1), and conventional GNN-based (96.1%, 0.95 F1) intrusion detection systems. Additional ablation studies confirm the contribution of entropy modeling, GNN embeddings, and RL-based mitigation to overall system effec tiveness. Future work will explore federated and distributed cyber defense architectures, integration with edge computing for IoT environments, and adaptive policy learning under large-scale network conditions, enabling real-time resilience against highly sophisticated adaptive cyber threats.