Real‐Time Risk‐Adaptive Access Control With <scp>DRCFM</scp> : A Scalable <scp>BERT</scp> ‐ <scp>LSTM</scp> ‐ <scp>GRU</scp> Framework for Secure Systems

ABSTRACT This study introduces a BERT‐LSTM‐GRU hybrid model for dynamic Risk‐Adaptive Access Control (RAdAC), designed to enhance predictive accuracy, robustness, and scalability in real‐time IoT and cloud environments. The proposed model integrates BERT's contextual embeddings with LSTM and GRU's sequential modeling capabilities, unified by a novel Dynamic Risk‐Aware Contextual Fusion Module (DRCFM) that employs risk‐aware attention and reinforcement learning to dynamically adapt to varying risk levels. Adversarial training and explainable AI techniques further bolster the model's resilience and transparency. Evaluated on five diverse datasets, the model achieves 96% accuracy, 97% F1 score, and 0.99 AUC‐ROC on the Human Activity Recognition dataset, outperforming state‐of‐the‐art baselines by up to 5% in accuracy. Under adversarial attacks (FGSM, PGD), it retains 92% accuracy (FGSM) and 90% accuracy (PGD) on the same dataset, demonstrating a 14% robustness improvement over baselines. Scalability analysis reveals a minimal 3% accuracy drop when scaling from 10 000 to 1 000 000 instances, with an inference time of 14 ms per instance, supporting real‐time applications. An ablation study underscores the contributions of each component, with BERT improving accuracy by approximately 5%–6%, LSTM by 5%, and GRU by 4%. These results position the model as a comprehensive solution for dynamic access control, with future research directions including multi‐modal data integration, edge deployment, and advanced explainability techniques to further enhance its applicability and trustworthiness in secure systems.