cs.CRMar 30, 2026

GMA-SAWGAN-GP: A Novel Data Generative Framework to Enhance IDS Detection Performance

Authors: Ziyu Mu, Xiyu Shi, Safak Dogan

Organizations: Institute for Digital Technologies, Loughborough University London, London E20 3BS, UK.

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Intrusion Detection System (IDS) is often calibrated to known attacks and generalizes poorly to unknown threats. This paper proposes GMA-SAWGAN-GP, a novel generative augmentation framework built on a Self-Attention-enhanced Wasserstein GAN with Gradient Penalty (WGAN-GP). The generator employs Gumbel-Softmax regularization to model discrete fields, while a Multilayer Perceptron (MLP)-based AutoEncoder acts as a manifold regularizer. A lightweight gating network adaptively balances adversarial and reconstruction losses via entropy regularization, improving stability and mitigating mode collapse. The self-attention mechanism enables the generator to capture both short- and long-range dependencies among features within each record while preserving categorical semantics through Gumbel-Softmax heads. Extensive experiments on NSL-KDD, UNSW-NB15, and CICIDS2017 using five representative IDS models demonstrate that GMA-SAWGAN-GP significantly improves detection performance on known attacks and enhances generalization to unknown attacks. Leave-One-Attack-type-Out (LOAO) evaluations using Area Under the Receiver Operating Characteristic (AUROC) and True Positive Rate at a 5 percent False Positive Rate confirm that IDS models trained on augmented datasets achieve higher robustness under unseen attack scenarios. Ablation studies validate the contribution of each component to performance gains. Compared with baseline models, the proposed framework improves binary classification accuracy by an average of 5.3 percent and multi-classification accuracy by 2.2 percent, while AUROC and True Positive Rate at a 5 percent False Positive Rate for unknown attacks increase by 3.9 percent and 4.8 percent, respectively, across the three datasets. Overall, GMA-SAWGAN-GP provides an effective approach to generative augmentation for mixed-type network traffic, improving IDS accuracy and resilience.

Figures & tables

Fig. 1: The architecture of GMA-SAWGAN-GP: Upper part shows the workflow of the proposed method, while the lower part illustrates the structures of the Generator, Critic, and AutoEncoder blocks. α,β\alpha,\beta are adaptive mixing weights output by the gating network fωf_{\omega} to balance LgAEL_{g}^{AE} and LdL_{d} .
  Number of samples
DatasetClassOriginal Training setAugmented Training setOriginal Test set
NSL-KDDNormal67343673439711
R2L995109952885
Probe11656216562421
DoS45927459277460
U2R521005267
TABLE I: Datasets Information
DatasetTraining DataIDS ModelAcc.(%)F1(%)NormalAbnormal
     Acc.(%)F1(%)Acc.(%)F1(%)
UNSW-NB15OriginalCNN81.8 ±\pm 4.768.8 ±\pm 4.489.1 ±\pm 6.588.8 ±\pm 3.348.1 ±\pm 8.548.7 ±\pm 6.2
DNN84.1 ±\pm 1.571.5 ±\pm 3.591.5 ±\pm 0.590.5 ±\pm 0.850.0 ±\pm 8.252.6 ±\pm 6.3
LSTM84.0 ±\pm 3.170.2 ±\pm 7.292.0 ±\pm 1.090.4 ±\pm 1.746.7 ±\pm 15.850.0 ±\pm 12.7
CNN-BiLSTM80.7 ±\pm 0.461.9 ±\pm 0.491.9 ±\pm 0.688.7 ±\pm 0.329.2 ±\pm 1.235.0 ±\pm 0.8
CNN-LSTM81.0 ±\pm 0.962.8 ±\pm 1.991.9 ±\pm 0.888.6 ±\pm 0.631.0 ±\pm 3.536.7 ±\pm 3.4
TABLE II: Binary Classification Comparison of Different IDS Models with three datasets for the Proposed model
DatasetTraining DataIDS ModelAcc.(%)F1(%)DoSReconnaissanceShellcodeWorms
     Acc. (%)F1(%)Acc.(%)F1(%)Acc.(%)F1(%)Acc.(%)F1(%)
UNSW-NB15OriginalCNN86.5 ±\pm 0.851.5 ±\pm 1.864.6 ±\pm 9.468.5 ±\pm 6.043.7 ±\pm 3.048.1 ±\pm 1.945.8 ±\pm 4.024.5 ±\pm 1.321.1 ±\pm 4.423.6 ±\pm 2.6
DNN82.5 ±\pm 1.343.9 ±\pm 3.429.2 ±\pm 13.638.4 ±\pm 13.744.6 ±\pm 1.742.3 ±\pm 0.937.9 ±\pm 4.627.2 ±\pm 2.424.6 ±\pm 2.621.0 ±\pm 3.1
LSTM84.2 ±\pm 1.240.8 ±\pm 2.058.3 ±\pm 13.858.6 ±\pm 8.233.1 ±\pm 1.836.4 ±\pm 1.814.4 ±\pm 3.017.4 ±\pm 2.10.0 ±\pm 0.00.0 ±\pm 0.0
CNN-BiLSTM81.1 ±\pm 0.536.1 ±\pm 1.719.6 ±\pm 1.426.2 ±\pm 1.533.0 ±\pm 1.839.7 ±\pm 1.918.7 ±\pm 5.316.8 ±\pm 1.64.8 ±\pm 4.68.5 ±\pm 7.8
CNN-LSTM82.6 ±\pm 0.544.8 ±\pm 1.145.7 ±\pm 1.848.2 ±\pm 1.730.2 ±\pm 0.340.1 ±\pm 0.832.7 ±\pm 0.719.9 ±\pm 0.916.4 ±\pm 4.325.6 ±\pm 5.7
TABLE III: Multi-Classification Results of UNSW-NB15 Dataset With Different IDS Models
DatasetTraining DataIDS ModelAcc.(%)F1(%)R2LProbeDoSU2R
     Acc.(%)F1(%)Acc.(%)F1(%)Acc.(%)F1(%)Acc.(%)F1(%)
NSL-KDDOriginalCNN76.8 ±\pm 0.356.6 ±\pm 1.41.8 ±\pm 0.63.5 ±\pm 1.177.7 ±\pm 2.878.5 ±\pm 1.780.0 ±\pm 1.587.6 ±\pm 0.921.1 ±\pm 4.233.8 ±\pm 5.6
DNN74.4 ±\pm 0.651.9 ±\pm 1.80.6 ±\pm 0.51.2 ±\pm 1.059.2 ±\pm 3.667.6 ±\pm 2.477.9 ±\pm 0.986.1 ±\pm 0.515.9 ±\pm 5.226.8 ±\pm 7.7
LSTM75.7 ±\pm 0.555.8 ±\pm 1.02.7 ±\pm 2.25.1 ±\pm 4.074.4 ±\pm 2.472.0 ±\pm 2.176.8 ±\pm 1.085.5 ±\pm 0.724.9 ±\pm 4.436.4 ±\pm 5.0
CNN-BiLSTM75.1 ±\pm 1.553.5 ±\pm 2.54.2 ±\pm 2.37.9 ±\pm 4.258.8 ±\pm 4.565.4 ±\pm 4.579.1 ±\pm 3.486.8 ±\pm 2.118.4 ±\pm 7.028.6 ±\pm 8.4
CNN-LSTM77.0 ±\pm 0.655.3 ±\pm 1.06.2 ±\pm 1.211.6 ±\pm 2.154.4 ±\pm 4.165.9 ±\pm 3.685.6 ±\pm 0.290.6 ±\pm 0.217.9 ±\pm 3.529.1 ±\pm 4.9
TABLE IV: Multi-Classification Results of NSL-KDD Dataset With Different IDS Models
DatasetTraining DataIDS ModelAcc.(%)F1(%)DoSPortScanBruteForceWebAttackBot
     Acc.(%)F1(%)Acc.(%)F1(%)Acc.(%)F1(%)Acc.(%)F1(%)Acc.(%)F1(%)
CICIDS2017OriginalCNN97.3 ±\pm 0.292.8 ±\pm 1.598.3 ±\pm 0.996.7 ±\pm 0.499.8 ±\pm 0.193.4 ±\pm 0.299.1 ±\pm 0.397.4 ±\pm 0.691.9 ±\pm 1.183.4 ±\pm 0.790.6 ±\pm 11.988.1 ±\pm 8.0
DNN96.8 ±\pm 0.289.3 ±\pm 2.593.9 ±\pm 2.595.5 ±\pm 0.698.3 ±\pm 1.893.2 ±\pm 0.398.6 ±\pm 0.797.6 ±\pm 0.485.7 ±\pm 1.080.7 ±\pm 1.160.5 ±\pm 16.671.1 ±\pm 13.6
LSTM96.5 ±\pm 0.390.0 ±\pm 0.897.0 ±\pm 1.195.8 ±\pm 0.399.3 ±\pm 0.293.0 ±\pm 0.198.0 ±\pm 2.192.0 ±\pm 2.788.7 ±\pm 1.081.9 ±\pm 0.680.4 ±\pm 3.379.4 ±\pm 2.6
CNN-BiLSTM98.1 ±\pm 0.294.9 ±\pm 0.598.9 ±\pm 0.398.0 ±\pm 0.397.6 ±\pm 1.594.1 ±\pm 0.699.0 ±\pm 0.498.6 ±\pm 0.393.9 ±\pm 1.284.5 ±\pm 1.898.8 ±\pm 0.395.5 ±\pm 0.9
CNN-LSTM98.4 ±\pm 0.095.5 ±\pm 0.199.5 ±\pm 0.198.5 ±\pm 0.097.1 ±\pm 0.794.6 ±\pm 0.199.2 ±\pm 0.199.1 ±\pm 0.094.3 ±\pm 0.485.4 ±\pm 0.598.8 ±\pm 0.296.9 ±\pm 0.4
TABLE V: Multi-Classification Results of CICIDS2017 Dataset With Different IDS Models
DatasetUnknown AttackModelOriginal DatasetAugmented Dataset
AUROC (%)TPR@5%FPR (%)AUROC (%)TPR@5%FPR (%)
NSL–KDDProbeCNN84.4 ±\pm 2.953.3 ±\pm 1.085.6 ±\pm 1.254.6 ±\pm 8.8
DNN68.3 ±\pm 11.551.4 ±\pm 4.775.7 ±\pm 9.472.5 ±\pm 13.2
LSTM72.5 ±\pm 4.744.9 ±\pm 11.076.4 ±\pm 2.539.5 ±\pm 10.3
CNN-BiLSTM81.3 ±\pm 4.453.5 ±\pm 12.283.1 ±\pm 2.056.3 ±\pm 10.1
CNN-LSTM78.4 ±\pm 4.131.0 ±\pm 12.079.5 ±\pm 3.433.5 ±\pm 11.1
TABLE VI: LOAO Results Across Five IDS Models on NSL-KDD, UNSW-NB15, and CICIDS2017
Fig. 2: PCA projections of real and synthetic traffic across three datasets. Rows label (a)–(c) correspond to CICIDS2017, UNSW-NB15, and NSL-KDD, respectively. Columns show the original training data and synthetic samples generated by different GAN-based models. In each plot, the horizontal and vertical axes correspond to the first and second principal components of PCA. Colors indicate a standardized scalar, with scales shown on the right side.
More in this topic

Synthetic Data Augmentation

Synthetic data augmentation creates or modifies training examples to improve learning when real data are scarce, imbalanced, costly, or sensitive.

See all 453 papers

Explore similar work

CardsList
Mar 19, 2026

A Novel Solution for Zero-Day Attack Detection in IDS using Self-Attention and Jensen-Shannon Divergence in WGAN-GP

The increasing sophistication of cyber threats, especially zero-day attacks, poses a significant challenge to cybersecurity. Zero-day attacks exploit unknown vulnerabilities, making them difficult to detect and defend against. Existing approaches patch flaws and deploy an Intrusion Detection System (IDS). Using advanced Wasserstein GANs with Gradient Penalty (WGAN-GP), this paper makes a novel proposition to synthesize network traffic that mimics zero-day patterns, enriching data diversity and improving IDS generalization. SA-WGAN-GP is first introduced, which adds a Self-Attention (SA) mechanism to capture long-range cross-feature dependencies by reshaping the feature vector into tokens after dense projections. A JS-WGAN-GP is then proposed, which adds a Jensen-Shannon (JS) divergence-based auxiliary discriminator that is trained with Binary Cross-Entropy (BCE), frozen during updates, and used to regularize the generator for smoother gradients and higher sample quality. Third, SA-JS-WGAN-GP is created by combining the SA mechanism with JS divergence, thereby enhancing the data generation ability of WGAN-GP. As data augmentation does not equate with true zero-day attack discovery, we emulate zero-day attacks via the leave-one-attack-type-out method on the NSL-KDD dataset for training all GANs and IDS models in the assessment of the effectiveness of the proposed solution. The evaluation results show that integrating SA and JS divergence into WGAN-GP yields superior IDS performance and more effective zero-day risk detection.

Ziyu Mu, Xiyu Shi, Safak Dogan

Institute for Digital Technologies, Loughborough University, London, UK

Network Intrusion DetectionGenerative Adversarial NetworksSynthetic Data Augmentation
AskPDF
Jul 1, 2026

Generative AI and Federated Learning for Intrusion Detection Systems: A Survey

Intrusion Detection Systems (IDSs) are essential for monitoring network traffic and identifying malicious activities in modern cyber-physical, Internet of Things (IoT), enterprise, and distributed network environments. However, developing reliable IDS models remains challenging because attack behaviors evolve over time, realistic datasets are difficult to obtain, traffic records may be incomplete, attack classes are often imbalanced, and privacy constraints limit centralized data collection. Recent advances in generative artificial intelligence (AI) and Federated Learning (FL) provide new opportunities to address these limitations. Generative models can support anomaly detection, synthetic traffic generation, data augmentation, data imputation, adversarial traffic generation, and IDS alert explanation. FL enables distributed IDS training without directly sharing local network traffic, making it suitable for privacy-sensitive and geographically distributed environments. This survey provides a structured review of generative AI and FL techniques for IDS. We first summarize representative IDS research directions, including adversarial machine learning, anomaly-based detection, IoT-oriented IDS, explainable IDS, and benchmark datasets. We then categorize generative AI applications in IDS according to model families and task objectives, covering autoencoder-based models, Generative Adversarial Networks (GANs), diffusion models, and Large Language Models (LLMs). Finally, we review emerging studies that integrate generative AI with FL-based IDS and discuss open challenges, including synthetic data quality, realistic traffic generation, dual-use adversarial risks, non-IID client distributions, communication-efficient model sharing, federated IDS benchmarking, and domain-specific LLMs for network security.

Jiefei Liu, Abu Saleh Md Tayeen, Pratyay Kumar +5

Department of Computer Science, New Mexico State University, Las Cruces, NM, USA · University of Hartford, CT, USA · DEVCOM Analysis Center, WSMR, NM, USA

Network Intrusion DetectionSynthetic Anomaly GenerationFederated Learning
AskPDF
Jul 17, 2026

Boundary-Seeking GAN-Augmented TabTransformer for Adversarially Robust Intrusion Detection

Machine learning-based intrusion detection systems (IDSs) often suffer from class imbalance and vulnerability to adversarial attacks, leading to degraded detection performance and reduced robustness. This study proposes a TabTransformer framework augmented by the Boundary-Seeking Generative Adversarial Network (BGAN) for flow-based intrusion detection using the CICIDS2017 dataset. BGAN serves a dual purpose by generating synthetic minority-class samples to mitigate data imbalance and producing adversarial samples to evaluate model robustness. Experimental results demonstrate that BGAN augmentation improves TabTransformer's Macro-F1 score from 82.96% to 86.50%, with the largest class-wise improvement observed for Web_Attack (F1 score: 0.29 to 0.61). Robustness evaluation shows that all non-augmented models experienced a 100% Performance Drop Rate (PDR) under adversarial testing, whereas all BGAN-augmented models achieved negative PDR values, indicating improved resilience. Furthermore, the augmented TabTransformer maintained stable and low False Triggered Rate (FTR) values (1.51%-2.92%) across all noise levels, compared with the BGAN-augmented Decision Tree, which reached 49.09% under benign perturbations. These findings demonstrate that BGAN consistently enhances both class balance and adversarial robustness, while the proposed BGAN-TabTransformer framework provides an effective and adaptive intrusion detection solution for adversarial network environments.

Raihan Sultan Pasha Basuki, Aliyah Kurniasih

Department of Computer Science, Universitas Ary Ginanjar, Menara 165, Cilandak No.kav 1 Lt 2 & 18, Jakarta Selatan, 12560, DKI Jakarta, Indonesia

Synthetic Data AugmentationTransformerAdversarial AttacksNetwork Intrusion DetectionAdversarial Robustness
AskPDF