Adversarial Attacks on Deep Learning-Based Intrusion Detection Systems on Challenges and Countermeasures
A. Jeyaram, A. Muthukumaravel · Advances in computational intelligence and robotics book series · 2025
DL-based IDS are crucial for detecting and reducing security risks in network settings. Nevertheless, these systems are susceptible to adversarial assaults that exploit model flaws. This research presents a defensive architecture that combines dynamic retraining, ensemble learning, real-time monitoring, and model interpretability to improve the resilience of IDS. Over subsequent years, the performance assessment reveals significant improvements in accuracy (from 0.85 to 0.94), precision (from 0.78 to 0.91), recall (from 0.89 to 0.95), and F1 score (from 0.83 to 0.93). Significantly, there was a reduction in false positive rates from 0.12 to 0.06 and a drop in false negative rates from 0.11 to 0.05. Analysis of feature significance serves to identify crucial aspects that influence the predictions made by a model, hence improving its interpretability. The suggested structure facilitates the ability of IDS to adjust and react to evolving threats efficiently.