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

Ziyu Mu; Xiyu Shi; Safak Dogan

cs.CRMar 19, 2026

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

Authors: Ziyu Mu, Xiyu Shi, Safak Dogan

Organizations: Institute for Digital Technologies, Loughborough University, London, UK

Abstract

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.

Figures & tables

Figure 1: The flowchart (left) shows the workflow of the enhanced performance evaluation process for IDS models, while the right side illustrates each step involved.

Figure 2: The structure of the proposed SA-JS-WGAN-GP. Loss_G, Loss_D, and Loss_C represent the loss values of G, D, and C, respectively.

Panel of Figure 3 — Figure 3: Distribution of data types of the training and testing dataset in the NSL-KDD dataset. (a) The original training dataset contains a significant proportion of Normal (53.46%) and DOS (36.46%) data, while other attack types, like R2L (0.8%) and U2R (0.04%), are rare. (b) Synthetic data has been generated and added to the binary classification training dataset, balancing the attack types. U2R now has a higher proportion (4.18%) compared to the original dataset. (c) The original NSL-KDD test dataset has proportions similar to the original training set. (d) Synthetic data has been generated and added to the multi-class training set, showing a more balanced distribution of attack types, likely helping the model better recognize each category.

	WGAN-GP	SA-WGAN-GP	JS-WGAN-GP	SA-JS-WGAN-GP
G	Dense layers: 128, 256, 512	Dense layers: 128, 256, 512 + SA layers	Dense layers: 64, 64, 128, 256, 512, 1024	Dense layers: 64, 64, 128, 256, 512, 1024, Dense layers in SA: 128, 256, 512
C	Dense layers: 512, 256, 128	Dense layers: 512, 256, 128 + SA layers	Dense layers: 128, 64, 32, 16, Output	Dense layers: 128, 64, 32, 16, Output, Dense layers in SA: 128 and 64
D	–	–	Dense layers: 128, 64, 32, 16, Output	Dense layers: 128, 64, 32, 16, Output, Dense layers in SA: 128 and 64
Learning Rate	G: $1\times 10^{-4}$ , C: $1\times 10^{-4}$	G: $1\times 10^{-4}$ , C: $1\times 10^{-4}$	G: $1\times 10^{-4}$ , C: $1\times 10^{-4}$ , D: $1\times 10^{-4}$	G: $1\times 10^{-4}$ , C: $1\times 10^{-4}$ , D: $1\times 10^{-4}$
Epochs	10,000	10,000	10,000	10,000
G Activation Function	LeakyReLU, Tanh	LeakyReLU, Tanh	LeakyReLU, Tanh	LeakyReLU, Tanh

Table 1: Architectures and Training Parameters of Various WGAN-GP Models

WGAN Variant	IDS Model	Acc.	Normal			Abnormal
WGAN Variant	IDS Model	Acc.	Rec.	Prec.	F1-score	Rec.	Prec.	F1-score
WGAN-GP	SVM	75.1 $\pm$ 0.0	88.1 $\pm$ 0.0	65.7 $\pm$ 0.0	75.3 $\pm$ 0.0	65.2 $\pm$ 0.0	87.9 $\pm$ 0.0	74.9 $\pm$ 0.0
	DT	78.8 $\pm$ 1.0	91.1 $\pm$ 1.8	70.4 $\pm$ 2.2	80.9 $\pm$ 1.3	49.5 $\pm$ 3.5	95.3 $\pm$ 1.6	80.3 $\pm$ 2.1
	DNN	75.7 $\pm$ 6.0	76.5 $\pm$ 9.1	83.1 $\pm$ 11.4	78.4 $\pm$ 2.0	76.5 $\pm$ 9.1	83.1 $\pm$ 11.4	78.4 $\pm$ 2.0
	CNN	76.6 $\pm$ 5.1	62.7 $\pm$ 10.6	96.5 $\pm$ 2.0	75.3 $\pm$ 7.5	96.7 $\pm$ 2.4	66.8 $\pm$ 5.9	78.8 $\pm$ 3.4
	LSTM	81.4 $\pm$ 0.7	97.4 $\pm$ 0.2	72.7 $\pm$ 2.4	82.1 $\pm$ 0.7	72.4 $\pm$ 3.4	96.9 $\pm$ 0.5	82.8 $\pm$ 2.1

Table 2: Binary Classification Comparison of Different IDS Models for Various WGAN-GP Variants

IDS Model	WGAN Variant	Acc.	Normal			Abnormal
IDS Model	WGAN Variant	Acc.	Rec.	Prec.	F1-score	Rec.	Prec.	F1-score
SVM	WGAN-GP	75.1 $\pm$ 0.0	88.1 $\pm$ 0.0	65.7 $\pm$ 0.0	75.3 $\pm$ 0.0	65.2 $\pm$ 0.0	87.9 $\pm$ 0.0	74.9 $\pm$ 0.0
	SA-WGAN-GP	73.3 $\pm$ 0.0	93.6 $\pm$ 0.0	62.7 $\pm$ 0.0	73.3 $\pm$ 0.0	57.9 $\pm$ 0.0	92.2 $\pm$ 0.0	71.2 $\pm$ 0.0
	JS-WGAN-GP	73.1 $\pm$ 0.0	98.2 $\pm$ 0.0	61.8 $\pm$ 0.0	73.1 $\pm$ 0.0	54.2 $\pm$ 0.0	97.5 $\pm$ 0.0	69.6 $\pm$ 0.0
	SA-JS-WGAN-GP	73.3 $\pm$ 0.0	98.2 $\pm$ 0.0	62.0 $\pm$ 0.0	73.3 $\pm$ 0.0	54.5 $\pm$ 0.0	97.5 $\pm$ 0.0	69.9 $\pm$ 0.0
DT	WGAN-GP	78.8 $\pm$ 1.0	91.1 $\pm$ 1.8	70.4 $\pm$ 2.2	80.9 $\pm$ 1.3	49.5 $\pm$ 3.5	95.3 $\pm$ 1.6	80.3 $\pm$ 2.1

Table 3: Binary Classification Comparison of the Proposed SA, JS, and SA-JS WGAN-GPs for Different IDS models

Panel of Figure 4 — Figure 4: Binary classification average accuracy of various IDS models with different WGAN-GP variants. (a) Average accuracy of IDS models across different WGAN-GP variants. (b) Average accuracy of WGAN-GP variants across different IDS models.

WGAN Variant	IDS Model	Acc.	DoS			Probe
WGAN Variant	IDS Model	Acc.	Rec.	Prec.	F1	Rec.	Prec.	F1
WGAN-GP	SVM	71.8 $\pm$ 0.0	70.5 $\pm$ 0.0	96.4 $\pm$ 0.0	81.5 $\pm$ 0.0	61.3 $\pm$ 0.0	69.4 $\pm$ 0.0	65.1 $\pm$ 0.0
	DT	73.4 $\pm$ 0.5	84.8 $\pm$ 2.1	95.6 $\pm$ 0.6	89.9 $\pm$ 1.2	72.8 $\pm$ 1.4	76.3 $\pm$ 0.4	74.5 $\pm$ 0.9
	DNN	77.7 $\pm$ 1.9	82.6 $\pm$ 2.5	90.5 $\pm$ 4.5	86.3 $\pm$ 1.3	74.2 $\pm$ 5.7	71.0 $\pm$ 2.6	72.5 $\pm$ 3.2
	CNN	79.1 $\pm$ 1.0	84.0 $\pm$ 1.5	91.8 $\pm$ 2.4	87.7 $\pm$ 1.5	82.9 $\pm$ 1.4	72.7 $\pm$ 3.7	77.4 $\pm$ 2.1
	LSTM	78.0 $\pm$ 5.2	80.7 $\pm$ 2.5	90.9 $\pm$ 3.9	85.5 $\pm$ 2.7	79.2 $\pm$ 7.9	72.0 $\pm$ 4.2	75.3 $\pm$ 5.5

Table 4: Multi-Classification Results of NSL-KDD Dataset with Different IDS Models for Different WGAN-GP Variants (Part 1: Acc., DoS and Probe)

WGAN Variant	IDS Model	R2L			U2R
WGAN Variant	IDS Model	Rec.	Prec.	F1	Rec.	Prec.	F1
WGAN-GP	SVM	-	-	-	-	-	-
	DT	14.6 $\pm$ 2.3	96.8 $\pm$ 1.1	25.3 $\pm$ 3.5	11.9 $\pm$ 0.0	30.0 $\pm$ 4.5	17.0 $\pm$ 0.9
	DNN	39.1 $\pm$ 15.7	71.6 $\pm$ 12.2	49.1 $\pm$ 5.5	26.6 $\pm$ 5.2	26.7 $\pm$ 13.2	24.7 $\pm$ 6.2
	CNN	9.4 $\pm$ 5.0	96.0 $\pm$ 2.8	16.8 $\pm$ 8.1	34.8 $\pm$ 2.9	23.4 $\pm$ 7.7	27.3 $\pm$ 5.5
	LSTM	50.9 $\pm$ 17.2	64.3 $\pm$ 5.4	54.0 $\pm$ 18.1	34.2 $\pm$ 4.7	6.7 $\pm$ 3.7	10.9 $\pm$ 5.3

Table 5: Multi-Classification Results of NSL-KDD Dataset with Different IDS Models for Different WGAN-GP Variants (Part 2: R2L and U2R)

IDS Model	WGAN Variant	Acc.	DoS			Probe
IDS Model	WGAN Variant	Acc.	Rec.	Prec.	F1	Rec.	Prec.	F1
SVM	WGAN-GP	71.8 $\pm$ 0.0	70.5 $\pm$ 0.0	96.4 $\pm$ 0.0	81.5 $\pm$ 0.0	61.3 $\pm$ 0.0	69.4 $\pm$ 0.0	65.1 $\pm$ 0.0
	SA-WGAN-GP	73.1 $\pm$ 0.0	71.1 $\pm$ 0.0	94.6 $\pm$ 0.0	81.2 $\pm$ 0.0	71.2 $\pm$ 0.0	78.9 $\pm$ 0.0	74.9 $\pm$ 0.0
	JS-WGAN-GP	74.4 $\pm$ 0.0	75.2 $\pm$ 0.0	97.2 $\pm$ 0.0	84.8 $\pm$ 0.0	69.1 $\pm$ 0.0	90.7 $\pm$ 0.0	69.1 $\pm$ 0.0
	SA-JS-WGAN-GP	74.6 $\pm$ 0.0	75.2 $\pm$ 0.0	97.2 $\pm$ 0.0	84.8 $\pm$ 0.0	69.1 $\pm$ 0.0	90.7 $\pm$ 0.0	69.1 $\pm$ 0.0
DT	WGAN-GP	73.4 $\pm$ 0.5	84.8 $\pm$ 2.1	95.6 $\pm$ 0.6	89.9 $\pm$ 1.2	72.8 $\pm$ 1.4	76.3 $\pm$ 0.4	74.5 $\pm$ 0.9

Table 6: Multi-Classification Results of NSL-KDD Dataset with Different WGAN-GP Variants for Different IDS Models (Part 1: Acc., DoS and Probe)

IDS Model	WGAN Variant	R2L			U2R
IDS Model	WGAN Variant	Rec.	Prec.	F1	Rec.	Prec.	F1
SVM	WGAN-GP	-	-	-	-	-	-
	SA-WGAN-GP	3.9 $\pm$ 0.0	84.3 $\pm$ 0.0	7.5 $\pm$ 0.0	10.4 $\pm$ 0.0	43.8 $\pm$ 0.0	16.9 $\pm$ 0.0
	JS-WGAN-GP	0.2 $\pm$ 0.0	40.0 $\pm$ 0.0	0.4 $\pm$ 0.0	6.0 $\pm$ 0.0	8.2 $\pm$ 0.0	6.9 $\pm$ 0.0
	SA-JS-WGAN-GP	0.2 $\pm$ 0.0	33.3 $\pm$ 0.0	0.5 $\pm$ 0.0	34.3 $\pm$ 0.0	63.9 $\pm$ 0.0	44.7 $\pm$ 0.0
DT	WGAN-GP	14.6 $\pm$ 2.3	96.8 $\pm$ 1.1	25.3 $\pm$ 3.5	11.9 $\pm$ 0.0	30.0 $\pm$ 4.5	17.0 $\pm$ 0.9

Table 7: Multi-Classification Results of NSL-KDD Dataset with Different WGAN-GP Variants for Different IDS Models (Part 2: R2L and U2R)

Panel of Figure 5 — Figure 5: Multi-classification average accuracy of various IDS models with different WGAN-GP variants. (a) Average accuracy of IDS models across different WGAN-GP variants; (b) Average accuracy of WGAN-GP variants across different IDS models

GAN Models	Acc. (B) (%)	Acc. (M) (%)	Training Time (hours)
SYN-GAN [ 30 ]	79.1 $\pm$ 5.6	85.1 $\pm$ 1.2	5.4
IGAN-IDS [ 17 ]	81.5 $\pm$ 4.2	79.8 $\pm$ 1.4	2.3
AI-based NIDS [ 27 ]	79.2 $\pm$ 7.8	79.8 $\pm$ 0.5	2.8
VAE-GAN [ 44 ]	83.3 $\pm$ 0.9	81.5 $\pm$ 3.1	3.9
Mts-dvGAN [ 38 ]	80.1 $\pm$ 1.6	80.7 $\pm$ 0.7	5
GANS-ECNN [ 36 ]	80.8 $\pm$ 3.6	76.1 $\pm$ 2.6	15

Table 8: Performance Comparison of the Proposed SA-, JS-, and SA-JS-WGAN-GP with the State-of-the-art GAN models on NSL-KDD Dataset Based on IDS Detection Accuracy for Binary Classification (B) and Multi-classification (M).

GAN model	IDS model	Acc	Macro-F1	AUROC	TPR@5%FPR
Baseline (No GAN)	LSTM	84.8 $\pm$ 1.1	71.7 $\pm$ 1.3	64.0 $\pm$ 7.2	7.0 $\pm$ 1.4
	CNN	85.1 $\pm$ 0.3	75.1 $\pm$ 1.1	72.7 $\pm$ 1.2	29.5 $\pm$ 8.7
	DNN	78.4 $\pm$ 6.9	61.4 $\pm$ 5.4	45.4 $\pm$ 8.8	14.4 $\pm$ 7.7
	DT	75.4 $\pm$ 17.6	56.4 $\pm$ 11.0	50.1 $\pm$ 0.1	0.2 $\pm$ 0.2
	SVM	80.8 $\pm$ 0.0	62.2 $\pm$ 0.0	47.0 $\pm$ 0.0	0.0 $\pm$ 0.0
WGAN-GP	LSTM	83.4 $\pm$ 1.4	60.8 $\pm$ 1.2	46.8 $\pm$ 11.0	2.9 $\pm$ 1.6

Table 9: Comparison of LOAO Results Across Different GAN and IDS Models in Emulation of Zero-day Attack Detection

Model	frac_lt_eps	Macro kNN p50	Macro kNN p95	Macro MMD 2 (clip 0)	ALL MMD 2
WGAN-GP	0.0	0.148	0.686	0.000512	0.023734
SA-WGAN-GP	0.0	0.309	1.312	0.002859	0.021832
JS-WGAN-GP	0.0	7.824	8.548	0.379318	0.285558
SA-JS-WGAN-GP	0.0	8.216	8.885	0.622680	0.343730

Table 10: kNN/MMD Alignment by Different WGAN-GP Variants (Five Classes)

Network Intrusion Detection

Network intrusion detection identifies malicious or abnormal network activity to help protect computers, connected devices, and critical infrastructure.

See all 168 papers

GMA-SAWGAN-GP: A Novel Data Generative Framework to Enhance IDS Detection Performance
Mar 30, 2026Ziyu Mu, Xiyu Shi, Safak DoganSynthetic Data Augmentation Network Intrusion Detection
Institute for Digital Technologies, Loughborough University London, London E20 3BS, UK.
ThreatFormer-IDS: Robust Transformer Intrusion Detection with Zero-Day Generalization and Explainable Attribution
Feb 26, 2026Srikumar NayakExplainable Intrusion Detection IoT Intrusion Detection
Incedo Inc, USA
Co-Evolving Zero-Day Jamming: Adaptive Attack Synthesis and Graph Attention-Based Online Detection
Sep 18, 2026Ghilas Aissou, Rémi A. Chou, Taejoon KimBlack-Box Adversarial Attacks Reinforcement Learning
Arizona State University, School of Electrical, Computer and Energy Engineering, Tempe, AZ 85287, USA · The University of Texas at Arlington, Department of Computer Science and Engineering, Arlington, TX 76010, USA

Mar 30, 2026J/K move · Enter open · S save

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

Ziyu Mu, Xiyu Shi, Safak Dogan

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

Synthetic Data Augmentation Network Intrusion Detection Generative Adversarial Networks

Open paper AskPDF arXiv

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.

Abstract

Network Intrusion Detection

Explore similar work

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

ThreatFormer-IDS: Robust Transformer Intrusion Detection with Zero-Day Generalization and Explainable Attribution

Co-Evolving Zero-Day Jamming: Adaptive Attack Synthesis and Graph Attention-Based Online Detection