AI-driven energy-efficient cybersecurity frameworks for sustainable digital infrastructures

• Traditional models focus on detection accuracy but lack energy efficiency, making them unsuitable for deployment in resource-constrained settings. • Introduction of a novel hybrid model: Dynamic Beetle Antennae Search–Mutated Adaptive Weighted Random Forest Tree (DynBAS-AWRF Tree). • Performs intelligent feature subset selection and hyperparameter tuning. • Uses labeled intrusion detection datasets with network flow and system behavior logs relevant to IoT/edge environments. • Preprocessing includes Min-Max normalization, one-hot encoding, and Linear Discriminant Analysis (LDA) for dimensionality reduction while preserving class separability. Cybersecurity in digital infrastructures faces energy limitations, especially in IoT systems. Traditional models emphasize accuracy but lack efficiency, making them unsuitable for edge deployment. Existing methods struggle to balance performance and energy use, particularly with high-dimensional, imbalanced data. The objective is to develop an energy-conscious, adaptive cybersecurity framework capable of delivering robust intrusion detection while minimizing computational overhead. A novel hybrid model named Dynamic Beetle Antennae Search–Mutated Adaptive Weighted Random Forest Tree (DynBAS-AWRF Tree) is introduced, integrating a dynamic feature selection and tuning mechanism with a highly adaptable, weighted ensemble classifier. The DynBAS component performs intelligent feature subset selection and hyperparameter tuning through directional sensing and mutation-enhanced search strategies. The AWRF Tree adapts its structure and weighting in response to class distribution and feature relevance, reducing complexity while improving sensitivity to minority attacks. Evaluation utilizes labeled intrusion detection datasets containing network flow and system behavior logs representative of smart IoT and edge environments. Min-max normalization and one-hot encoding are applied to standardize and structure the input data for optimal model ingestion. Linear Discriminant Analysis (LDA) reduces dimensionality while preserving class separability. DynBAS selects optimal feature sets and hyperparameters, while AWRF Tree dynamically adjusts node weights and tree depth, maintaining efficiency and adaptability. The hybrid model achieves competitive detection accuracy with significantly reduced feature usage by 45 % and increased energy efficiency from 50 % to 85 %. The DynBAS-AWRF Tree framework enables a scalable and sustainable cybersecurity solution suitable for deployment in constrained digital infrastructures.