Answers are grounded in this paper and cite their sources.
Modern society increasingly relies on online applications, which demand high-quality networking services. Network Intrusion Detection Systems (NIDS) are vital for detecting and mitigating cyberattacks that aim to disrupt such services. While the research community continues to develop innovative NIDS, several limitations persist in state-of-the-art systems, including a heavy reliance on supervised configuration, burdensome hyperparameter tuning, and a failure to account for traffic contamination in training data. In this work, we address these challenges by proposing an unsupervised hybrid NIDS that integrates Generative Adversarial Network traffic modeling with Dynamic Dictionaries of legitimate and suspicious IP Addresses (GAN-DDA). Specifically, a fast Anomaly GAN (f-AnoGAN) is designed to reconstruct historical IP flow traffic and attribute higher residuals to anomalous samples. An adaptive threshold is integrated into the model for autonomous traffic classification, demonstrating robustness against low levels of traffic contamination in the training sets. A post-processing stage further refines inferences based on the state of dynamic dictionaries. The proposed system achieves F1-scores of 0.9215, 0.9649, and 0.9041 on the UNSW-NB15, CIC-IDS2017, and CSE-CIC-IDS2018 benchmarks, respectively. An ablation study validates the necessity of each internal component for achieving this performance. Comparisons with state-of-the-art alternatives indicate competitive performance, though differences in experimental setups across studies highlight the need for standardized future evaluations. Finally, future research directions are identified to improve the statistical generalization, long-term performance, and applicability of reconstruction-based NIDS to real networks.
No takes yet. Share an insight, caveat, or question.
Ruffo et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: