Adversarial Attack Detection in Network Traffic Using Bi-LSTM and Dynamic Feature Selection
Yiwei Su · Advances in transdisciplinary engineering · 2025
The rapid development of interconnected systems has led to an increase in the incidence of adversarial attacks, highlighting the necessity of interdisciplinary approaches to tackle these challenges effectively. These attacks use malicious perturbations embedded in network traffic to avoid detection. Traditional intrusion detection systems, such as rule-based methods and statistical models, often fail to cope with such dynamic and complex attack techniques due to their inability to adapt to the nonlinear, high-dimensional, and time-varying characteristics of network data. This study proposes a novel framework for adversarial attack identification that integrates advancements from multiple disciplines, including machine learning, data science, and network security. By leveraging the attention mechanism, dynamic feature selection, and bi-directional long and short-term memory (Bi-LSTM) networks, the framework addresses these challenges comprehensively. The Bi-LSTM model processes bi-directional time-dependent network traffic, incorporating insights from temporal data analysis. Meanwhile, the attention mechanism, inspired by cognitive science applications in artificial intelligence, identifies and prioritizes critical features, enhancing detection accuracy and interpretability. Additionally, the dynamic feature selection algorithm, rooted in statistical modeling and data preprocessing, improves robustness and reliability by eliminating irrelevant and noisy features. This interdisciplinary approach enables the framework to identify DoS, Probe, R2L, and U2R attack types using benchmark datasets such as KDD Cup 1999 and NSL-KDD with up to 96% accuracy, demonstrating the critical role of cross-disciplinary applications in advancing intrusion detection systems against sophisticated adversarial attacks.