End-to-End Network Security Prediction via Dual Attention- Enhanced Situation Assessment and IHHO-GResNeSt Integration

Abstract

To address the limitations of existing cybersecurity prediction methods in terms of insufficient feature extraction and low parameter optimization efficiency, this study proposes an end-to-end cybersecurity prediction method based on Channel-Spatial Dual Attention (CS-DA) to enhance situation assessment and integrating an improved Harris Hawks Optimization algorithm and a Global Context Residual Split-Attention Network (IHHO-GResNeSt). This method first introduces CS-DA to enhance the extraction of key threat features during the cybersecurity status assessment stage, and then uses IHHO to globally optimize the assessment model parameters. Subsequently, quantified situation values are used as prior input, and GResNeSt is introduced for temporal feature learning. The Sparrow Search Algorithm (SSA) is then combined with the GResNeSt algorithm to optimize the prediction hyperparameters, achieving integrated end-to-end security situation prediction that combines assessment and prediction. Experimental results on three public benchmark datasets-NSL-KDD, UNSW- NB15, and CICIDS2017 show that the proposed method achieves a false negative rate of 0.05 and a false positive rate of 0.03 in the security assessment stage, with a situational assessment bias of 0.04 and an inference latency of only 32.16 ms, demonstrating good assessment accuracy and real-time performance. In the prediction stage, the RMSE (0.08) of the SSA-tuned model is superior to CNN- LSTM (0.17), XGBoost (0.20), and Transformer (0.15). These results demonstrate that the proposed method can achieve high-precision, robust, and efficient cybersecurity prediction in complex and dynamic network environments, providing reliable support for security early warning and decision- making.