ACGAN-GP-Based synthetic data generation to handle imbalanced data in intrusion detection systems

Obtaining adequate representations of each class is crucial for accurately identifying a wide range of attacks using machine learning (ML) and deep learning (DL) approaches. However, in real-world applications, intrusion detection systems (IDS) are biased toward the majority classes, and datasets are highly skewed. SMOTE, undersampling, and oversampling are popular data augmentation techniques, but they only slightly improve detection performance. This study proposes an autoencoder for feature extraction, an Auxiliary Classifier Generative Adversarial Network (ACGAN) for generating synthetic data to address class imbalance, and Logistic Regression, KNN, Random Forest, Gradient Boosting, Extra Trees, XGBoost, LightGBM, and HGBoost classifiers for classifying the balanced dataset. A gradient penalty is used in ACGAN to ensure consistent model training. The publicly available and very popular benchmark dataset, NSL-KDD, is used to test the model. The framework is also assessed on UNSW_NB15 to examine performance across datasets. To evaluate the performance of the autoencoder and ACGAN-GP, this paper uses the loss curve, reconstructed error distribution, and latent feature representation, as well as precision, recall, F1-score, accuracy, ROC curve, and Wilcoxon model, to assess the outcomes of the selected classifiers. The training and validation losses for the autoencoder decreased sharply, and it accurately extracted complex features. The ACGAN maintains training consistency and successfully produces high-quality synthetic samples to solve the data imbalance problem. Among ensemble models, the Extra Trees classifier achieves the highest accuracy, 99.60%, followed by Random Forest at 99.55%. Among traditional ML models, KNN achieves the highest accuracy, 99.40%, and LR achieves slightly lower accuracy, 94.62%. Among boosting models, LightGBM achieves the highest accuracy, 99.60%, followed by XGBoost at 99.59%. Among all the classifiers, Extra Trees and LightGBM perform well, with equal accuracy of 99.60%, while XGBoost is close (99.59%). Based on the Macro-AUC value, XGBoost is 0.998, followed by LightGBM (0.997). In terms of statistical analysis, Extra Trees performs well compared to other classifiers. The NSL_KDD dataset's performance is significantly better than that of the UNSW_NB15 dataset. From these results, it is clear that the proposed approach successfully generates synthetic data to address the data imbalance problem, and the classification results support the model's performance.

Authors

Institutions

Publication Details

Journal
PLoS ONE
Published
2026-11-20
DOI
https://doi.org/10.1371/journal.pone.0362836
Primary Topic
Network Security and Intrusion Detection
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

ACGAN-GP-Based synthetic data generation to handle imbalanced data in intrusion detection systems

Pintu Chandra Shill, Md Tariqul Islam, Sohel Rana, Md Shohag Hasan
PLoS ONE
Network Security and Intrusion Detection
article

ACGAN-GP-Based synthetic data generation to handle imbalanced data in intrusion detection systems

Pintu Chandra Shill, Md Tariqul Islam, Sohel Rana, Md Shohag Hasan
article en

Abstract

Obtaining adequate representations of each class is crucial for accurately identifying a wide range of attacks using machine learning (ML) and deep learning (DL) approaches. However, in real-world applications, intrusion detection systems (IDS) are biased toward the majority classes, and datasets are highly skewed. SMOTE, undersampling, and oversampling are popular data augmentation techniques, but they only slightly improve detection performance. This study proposes an autoencoder for feature extraction, an Auxiliary Classifier Generative Adversarial Network (ACGAN) for generating synthetic data to address class imbalance, and Logistic Regression, KNN, Random Forest, Gradient Boosting, Extra Trees, XGBoost, LightGBM, and HGBoost classifiers for classifying the balanced dataset. A gradient penalty is used in ACGAN to ensure consistent model training. The publicly available and very popular benchmark dataset, NSL-KDD, is used to test the model. The framework is also assessed on UNSW_NB15 to examine performance across datasets. To evaluate the performance of the autoencoder and ACGAN-GP, this paper uses the loss curve, reconstructed error distribution, and latent feature representation, as well as precision, recall, F1-score, accuracy, ROC curve, and Wilcoxon model, to assess the outcomes of the selected classifiers. The training and validation losses for the autoencoder decreased sharply, and it accurately extracted complex features. The ACGAN maintains training consistency and successfully produces high-quality synthetic samples to solve the data imbalance problem. Among ensemble models, the Extra Trees classifier achieves the highest accuracy, 99.60%, followed by Random Forest at 99.55%. Among traditional ML models, KNN achieves the highest accuracy, 99.40%, and LR achieves slightly lower accuracy, 94.62%. Among boosting models, LightGBM achieves the highest accuracy, 99.60%, followed by XGBoost at 99.59%. Among all the classifiers, Extra Trees and LightGBM perform well, with equal accuracy of 99.60%, while XGBoost is close (99.59%). Based on the Macro-AUC value, XGBoost is 0.998, followed by LightGBM (0.997). In terms of statistical analysis, Extra Trees performs well compared to other classifiers. The NSL_KDD dataset's performance is significantly better than that of the UNSW_NB15 dataset. From these results, it is clear that the proposed approach successfully generates synthetic data to address the data imbalance problem, and the classification results support the model's performance.

PLoS ONEVol. 21(11)
International Islamic University Chittagong (BD), East West University (BD)
Network Security and Intrusion Detection
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon