Network traffic anomaly detection faces critical challenges in feature extraction robustness, and computational efficiency due to increasing data dimensionality and environmental noise. Existing deep learning approaches offer partial improvements but suffer from noise sensitivity, structural information neglect, and unnecessary computational overhead. This paper presents an intelligent approach integrating attention mechanisms with deep feature learning, combining multi-scale attention dynamics and optimised gradient boosting to address network anomaly detection challenges. The core contributions encompass a hybrid solution that achieves noise-resilient feature extraction through self-attention weighting while preserving structural traffic patterns, coupled with an enhanced and optimised gradient boosted decision tree classifier employing logarithmic loss optimisation and early stopping mechanisms for effective high-dimensional sparse data processing. Experimental results demonstrate that the proposed approach outperforms the state-of-the-art baselines, including superior detection accuracy 98.2% and around 34.7% detection time reduction.
Guihua Wu (Thu,) studied this question.