An Effective Feature Selection‐Based Cyber Attack Detection Using a Polymorphic Graph Gudermannian Neural Network With Encryption Approach

ABSTRACT The rapid expansion of digitalization has intensified cybersecurity risks, exposing critical network vulnerabilities despite significant advances in encryption and intrusion detection systems (IDS). Many existing deep learning–based IDS still struggle with high false‐positive rates, misclassification, and limited adaptability, reducing their effectiveness in real‐time defense scenarios. To address these limitations, this study proposes a Polymorphic Graph Gudermannian Neural Network integrated with Adaptive Chaotic Satin Bowerbird Optimization (PG‐GNN‐AC‐SBO), complemented by a lightweight encryption mechanism. The framework incorporates a Fuzzy K‐Top Matching Value (FKTMV) module for robust preprocessing and normalization, along with a Hybrid Cat Hunting Sea‐Horse Optimizer (H‐CHO‐SHO) for efficient and interpretable feature selection. The PG‐GNN classifier employs graph‐based learning and a Gudermannian nonlinear activation function to effectively capture complex traffic behavior, while AC‐SBO dynamically tunes hyperparameters to enhance stability and classification accuracy. To ensure data confidentiality, a Synchronously Scrambled Diffuse Encryption (SSDE) scheme is applied, delivering strong security with low computational overhead. Experimental evaluations on the NSL‐KDD and CICIDS2017 datasets demonstrate the superiority of the proposed approach, achieving up to 99.82% accuracy and outperforming state‐of‐the‐art methods. The encryption and decryption times of 3.50 and 3.55 ms further confirm the model's lightweight design. Overall, the proposed system provides high throughput with minimal latency, demonstrating strong potential for real‐time and large‐scale cybersecurity deployments.