The emergence of Generative Adversarial Network (GAN) techniques has garnered significant attention from the research community for the development of Intrusion Detection Systems (IDS). However, conventional GAN-based IDS models face several challenges, including training instability, high computational costs, and system failures. To address these limitations, we propose a Hybrid Wasserstein GAN and Autoencoder Model (WGAN-AE) for intrusion detection. The proposed framework leverages the stability of WGAN and the feature extraction capabilities of the Autoencoder Model. The model was trained and evaluated using two recent benchmark datasets, 5GNIDD and IDSIoT2024. When trained on the 5GNIDD dataset, the model achieved an average area under the precision-recall curve is 99.8% using five-fold cross-validation and demonstrated a high detection accuracy of 97.35% when tested on independent test data. Additionally, the model is well-suited for deployment on resource-limited Internet-of-Things (IoT) devices due to its ability to detect attacks within microseconds and its small memory footprint of 60.24 kB. Similarly, when trained on the IDSIoT2024 dataset, the model achieved an average PR-AUC of 94.09% and an attack detection accuracy of 97.35% on independent test data, with a memory requirement of 61.84 kB. Extensive simulation results demonstrate that the proposed hybrid model effectively addresses the shortcomings of traditional GAN-based IDS approaches in terms of detection accuracy, computational efficiency, and applicability to real-world IoT environments.
Autoencodercybersecuritygenerative adversarial networkInternet of Thingsintrusion detection systemKing Khalid UniversityRGP.2/245/46Princess Nourah bint Abdulrahman UniversityPNURSP2025R760Deanship of Graduate Studies and Scientific Research at Najran UniversityNU/GP/SERC/13/352-1Introduction
Intrusion detection systems (IDSs) play a very important role in protecting IoT networks from a number of cyberattacks that are mainly targeted at device and network vulnerabilities such as the use of insecure communication protocols or weak authentication mechanisms [1]. Most of these attacks are launched through malicious software or firmware updates that may result in unauthorized access to the network or control of IoT devices [2]. Cyberattacks can violate the confidentiality, integrity, and availability of the network and may result in data breaches, unauthorized access to sensitive information, and denial of service [3,4]. The main challenge in this regard is the resource-constrained nature of IoT devices that pose a challenge to implementing advanced and robust security frameworks. The increasing complexity of security algorithms results in system outages, reduced performance, or poor service quality in IoT networks [5].
Over the past few years, GANs have received great attention because of their ability to learn data distribution and generate synthetic data that can mimic possible attack patterns [6]. Traditional signature-based approaches are also ineffective in identifying new and unexpected cyberattacks, whereas GANs learn to generate adversarial examples to identify outliers [7]. However, there are several shortcomings of the traditional GAN variants as well. For example, vanilla GAN [8] has some drawbacks such as model collapse and training instability that limit its effectiveness in generating diverse and representative samples. Conditional GANs [9] are more context-sensitive than their counterparts but they need labeled data. CycleGANs [10] are very efficient in the domain translation task, even in the absence of paired data, but they are usually slow and not very accurate in identifying small anomalies. The Wasserstein GAN [11] also improves the stability and diversity but at the expense of increasing the model complexity.
To overcome these issues, this paper proposes a Wasserstein GAN with Autoencoders (WGAN-AE) that integrates the best aspects of WGANs and autoencoders to develop a more efficient and effective intrusion detection system for IoT networks. The WGAN-AE can produce stable and diverse samples by leveraging the Wasserstein distance, which measures the difference between two probability distributions. Unlike traditional distance metrics, it provides a smoother and more meaningful way to compare real and generated data, leading to better training stability and improved detection of various attack types [12]. Autoencoders assist in detecting the changes that are likely to be a sign of a cyberattack by learning the detailed patterns of the normal IoT traffic [13]. This approach is less dependent on large datasets suitable for dynamic IoT environments. Furthermore, the WGAN-AE offers a complete solution that entails the reconstruction and anomaly identification of any input by comparing it with the learned normal behavior. This makes training more stable and enhances the novel attack detection rate that may not be detected by other models. A comparison of the proposed WGAN-AE with state-of-the-art GAN variants is presented in Table 1. The major contributions of the article are summarized as follows.
A comparison of state-of-the-art GAN variants with the proposed WGAN-AE in the context of IDS
Feature
GAN variants
Vanilla GAN
Conditional GAN
Cycle GAN
Wasserstein GAN
WGAN-AE
Training stability
Low
Moderate
Moderate
High
Very high
Mode collapse
High
Moderate
Moderate
Low
Very low
Feature learning
Limited
Good
Good
Strong
Strong
Anomaly detection
Poor
Moderate
Good
Good
Excellent
Training complexity
Moderate
High
High
Moderate
Moderate
Robustness
Low
Moderate
Moderate
High
Very high
A hybrid intrusion detection framework, WGAN-AE, is proposed by utilizing the key strengths of WGAN and AE. The model enhances the detection of emerging attack vectors by using Wasserstein distance for more stable training and autoencoder-based feature extraction.
The research introduces an unsupervised GAN-based approach that learns the normal operation of the system and identifies changes as possible cyberattacks. It also achieves this by properly distinguishing between the normal and attack behaviors through the use of adversarial training of the generator to generate diverse attack types.
A comprehensive evaluation framework is developed to assess the computational efficiency and effectiveness of the proposed scheme by using two advanced and comprehensive IDS benchmark datasets.
The remainder of the article is organized as follows. In Section 2, we present some latest contributions related to GAN-based IDS frameworks and describe some preliminaries. In Section 3, we elaborate on the research methodology and the design of the proposed framework. In Section 4, we present a brief discussion of the experimental setup and outcomes. Finally, in Section 5, we conclude the research with future research directions.
Related Work
This section overviews some significant contributions related to GAN-based IDS architectures. Rahman et al. [6] explored the potential of GAN for intrusion detection in IoT. The proposed scheme significantly decreased the dependency on real-world data. To analyze the effectiveness of designed model, the authors conducted extensive experiments on three open-source datasets including UNSW-NB15, NSL-KDD and BoT-IoT datasets. Message Queuing Telemetry Transport (MQTT) is a widely adapted network protocol in IoT infrastructures because of its lightweight and flexible nature. Boppana and Bagade [14] proposed a novel unsupervised GAN and autoencoder-based model GAN-AE, to detect unknown intrusions in MQTT-based IoT applications. The experimental results demonstrate the effectiveness of the proposed method against various modern IDS approaches. In another study, Li et al. [15] proposed a hybrid IDS model to detect Denial of Service (DoS)/botnet attacks in IoT systems. The authors designed an anomaly-based detection model called CL-GAN (CNN-LSTM GNN), combining a Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) with GAN to define a baseline of normal activity and identify malicious traffic. The proposed architecture was evaluated with NSL-KDD, CICIDS2018, and Bot-IoT datasets.
de Araujo-Filho et al. [16] presented a novel fog-based unsupervised IDS using GANs. The proposed IDS was developed for a fog architecture to meet the low-latency requirements of cyberphysical systems. Experimental outcomes indicated the higher detection rates and superior performance of the proposed approach over baseline models using three datasets. Zeghida et al. [17] proposed GAN-based methods to achieve a balanced dataset for greater attack detection accuracy. Additionally, they introduced three dedicated IDSs for attack detection using the MQTT protocol based on hybrid deep learning (DL) algorithms: Convolutional Neural Network with Recurrent Neural Network (CNN-RNN), CNN with Long Short-Term Memory (CNN-LSTM), and CNN with Gated Recurrent Unit (CNN-GRU). The experimental results demonstrated that the generated dataset had a superior performance in multiclass configuration. In another study, Das et al. [18] proposed two models: a Feedforward Neural Network (FNN) network and a CNN. The proposed models have been trained and tested on standard datasets as well as synthetic datasets. The generation of this synthetic data employs a Conditional Tabular Generative Adversarial Network (CTGAN). The experimental outcomes indicated less training time and memory utilization than several baseline models.
Wang et al. [19] proposed a Multi-Critics GAN to address the data imbalance issues in IDS systems. The authors analyzed the generated data quality by using Principal Component Analysis (PCA) plots and correlation heatmaps. Subsequently, they incorporated a hybrid CNN-LSTM model to analyze the clusters to achieve the promising performance of IDS systems. The experimental results confirmed the higher attack detection rate with better generalization. Dong et al. [20] presented a novel IDS framework MasqueradeGAN-GP (Generative Adversarial Networks with Gradient Penalty), for 6G networks by integrating a WGAN with a Gradient Penalty. In the proposed architecture, a generator transforms the anomalous traffic into a semblance of benign activity and the discriminator discerns the genuine and adversarial traffic. The efficacy of the designed models was investigated by conducting extensive experiments using two open-source datasets. Brabin et al. [21] presented a Cycle-Consistent GAN-based attack detection and secure data transmission framework for smart cities. The designed framework incorporated a Wild horse optimizer for feature extraction. Subsequently, the selected features are provided to the cycle-consistent GAN classifier to distinguish normal and malicious traffic. Furthermore, to ensure a secure data transmission, the authors incorporated Advanced Encryption Standard (AES) with Chameleon Swarm Algorithm.
The aforementioned works present a significant contribution towards GAN-based IDSs. However, the existing studies have a few shortcomings, including reliance on outdated datasets that fail to capture real-time IoT security challenges and a primary focus on attack detection accuracy while neglecting critical constraints such as memory requirements and computational efficiency. While some approaches address data imbalance using advanced GAN variants like Multi-Critics GAN, CTGAN, and WGAN with Gradient Penalty, others integrate federated learning for enhanced security in Fog-assisted IoT networks. However, these studies lack a holistic evaluation framework. To overcome these shortcomings, this article proposes a WGAN-AE, which is trained and evaluated using two real-time benchmark datasets. The proposed approach ensures a comprehensive performance analysis, including accuracy, memory consumption, system latency, and cross-validation, making it a more robust solution for modern IoT-based IDS.
Research Methodology and the Proposed Framework
This section briefly describes the dataset description, the design of the proposed architecture, and the training process. The workflow of the proposed architecture is presented in Fig. 1.
Block diagram of the proposed architecture
Dataset Description
To train and evaluate the proposed architecture’s performance, we utilized three of the latest and most comprehensive datasets. The following provides a detailed description.
5G-NIDD Dataset
The 5G-NIDD dataset is a fully labeled collection of network traffic data generated from a functional 5G test network at the University of Oulu, Finland [22]. The dataset includes various attack scenarios, such as Denial of Service (DoS) attacks, Internet Control Message Protocol (ICMP) Flood, synchronize (SYN) Flood, User Datagram Protocol (UDP) Flood, Hypertext Transfer Protocol (HTTP) Flood, and Slowrate and port scans, including Transmission Control Protocol (TCP) Connect Scan, SYN Scan, and UDP Scan. The detailed distribution of the 5G-NIDD dataset is presented in Fig. 2.
Distribution of 5G-NIDD dataset
IDSIoT2024 Dataset
The IDSIoT2024 dataset [23] is a comprehensive, real-time collection of network traffic data from an Internet of Things (IoT) environment comprising seven diverse smart devices: a smartwatch, surveillance camera, smartphone, laptop, smart vacuum, smart TV, and smart light. In this setup, the laptop serves a dual role: continuously monitoring and logging network traffic for analysis and actively executing various network-based attacks to simulate potential security threats. The dataset includes seven main categories: DoS, Injection, Man-in-the-Middle (MITM), malware, normal, routing, and vulnerability analysis. The detailed distribution of the IDSIoT2024 dataset is presented in Fig. 3.
Distribution of IDSIoT2024 dataset
The Proposed Architecture
This section provides a detailed description of each module of the proposed architecture.
Data Preprocessing
The preprocessing of the dataset 𝒟={(xi,yi)}i=1N is an essential step to prepare both the input featuresX and target labels y for the hybrid WGAN-AE and autoencoder model. This stage typically involves scaling the features and encoding the categorical target labels for classification tasks.
Feature Scaling (Standardization): The first step is to normalize the feature data so that each feature has zero mean and unit variance. This ensures that all features contribute equally during training, preventing features with larger ranges from dominating the optimization process.
Given a feature matrix X∈RN×d, where N is the number of samples and d is the number of features, each feature column xj∈RN is standardized using z -score normalization. The transformation is performed as follows (1):
xjscaled=xj−μjσj,where:
μj=1N∑i=1Nxij is the mean of the j-th feature across all samples.
σj=1N∑i=1N(xij−μj)2 is the standard deviation of the j-th feature.
This process ensures that the scaled feature xjscaled has (2):
E[xjscaled]=0 and Var(xjscaled)=1.
The entire feature matrix Xscaled is then used as the input for the model.
Label Encoding: In classification tasks, the target labels are often categorical, so we need to convert the labels into a numerical form that can be used by machine learning models. Label encoding is a common method for transforming categorical labels into integers.
Let yi∈{1,2,…,C} be the categorical class label of the i-th sample, where C is the number of classes. The goal is to map each class label to an integer, so the transformation is defined as (3):
yiencoded=LabelEncoder(yi),where yiencoded∈{0,1,…,C−1} represents the corresponding index of the label in a sorted list of unique classes.
The resulting encoded labels are stored in a vector yencoded∈RN, where each yiencoded represents the integer value corresponding to the original class label.
One-Hot Encoding: While label encoding is a simple way of converting categorical labels into integers, for many classification tasks, especially in neural networks, it is more effective to represent each class label as a one-hot encoded vector. In one-hot encoding, each class label is represented as a vector where only the index corresponding to the label is 1, and all other indices are 0.
For a given encoded label yiencoded∈{0,1,…,C−1}, we convert it into a one-hot vector yionehot∈RC. The one-hot encoding is given by (4):
yionehot=[0,0,…,1,…,0]where the 1 appears at the index yiencoded, and all other positions are 0.
The entire matrix of one-hot encoded labels is represented as Yonehot∈RN×C, where each row yionehot corresponds to the one-hot encoded vector for the i-th sample.
Hybrid WGAN-AE Model
The WGAN-AE is an upgraded model that combines the best features of autoencoders with GANs to offer superior attack detection in IoT networks. The model has two main modules: Autoencoder and Discriminator. In addition to that, the Wasserstein loss helps improve the training stability since the overall model learned to reconstruct the input data and distinguish between real and that which has been simulated. Mathematical formulation of the Hybrid WGAN-AE model is presented in this section.
A. Generator Architecture (Autoencoder)
A combination of an encoder and a decoder module constitutes the generator using the autoencoder structure. The autoencoder learns to accurately compress/input data into a low-dimensional representation, able to reconstruct the same input data from low-dimensionality space.
Encoder: The encoder maps the input data xi from the input space Rd to a lower-dimensional latent space. The encoder can be defined by a function fθ:Rd→Rz, where θ represents the parameters of the encoder, and z∈Rz is the latent representation. Mathematically, the encoder can be expressed as:
zi=fθ(xi),where zi is the compressed representation of xi.
Decoder: The decoder is responsible for reconstructing the original input xi from the latent vector zi. The decoder function gϕ:Rz→Rd is parameterized by ϕ, and the reconstruction x^i is given by:
x^i=gϕ(zi),where x^i is the reconstructed input.
The loss function for the autoencoder is typically the reconstruction loss, which measures how well the decoder can approximate the original input xi from the latent code zi. The reconstruction loss is given by:
Lrecon(xi,x^i)=‖xi−x^i‖22,where ‖⋅‖22 denotes the squared Euclidean distance.
B. Discriminator Architecture
The discriminator is a neural network that distinguishes between real data (from the dataset) and fake data (generated by the autoencoder). It is defined as a binary classifier Dψ:Rd→[0,1], where Dψ(xi)=1 indicates that xi is real and Dψ(xi)=0 indicates that xi is fake.
For the input data xi, the discriminator outputs a scalar Dψ(xi) representing the probability that xi is real. The discriminator is trained to maximize this probability when the data is real and minimize it when the data is fake. Thus, the loss for the discriminator can be described as (5):
Ldisc(xi,x^i)=−Ex∼Preal[logDψ(x)]−Ex^∼Pgen[log(1−Dψ(x^))],where Preal indicates the distribution of real data and Pgen represents the distribution of generated data.
The goal of the discriminator is to correctly classify real data as 1 and fake data as 0, minimizing the above loss function.
C. Wasserstein Loss for WGAN-AE
A key component of the Hybrid WGAN-AE is the Wasserstein loss, which is used to stabilize the training process, and to generate better samples.
The Wasserstein loss [24] for the discriminator is given by (6):
LWasserstein(xi,x^i)=Ex∼Preal[Dψ(x)]−Ex^∼Pgen[Dψ(x^)].
This loss function trains the discriminator to give high values to real data and low values to the generated data, thus trying to increase the distance between the two distributions. The Wasserstein loss has smoother gradients and solves the vanishing gradient problem that is typical for standard GANs.
The generator is trained to minimize the Wasserstein loss in the opposite direction, i.e., to minimize −LWasserstein, which ensures that the generated data distribution approaches the real data distribution.
D. Combined Model
In the Hybrid WGAN-AE, the combined model is trained alongside both the generator and the discriminator. The purpose of the combined model is to train the generator to create data that looks realistic enough to fool the discriminator. Nevertheless, the generator tries to minimize the reconstruction error as well so that the produced data is both realistic and consistent with the input data.
The total loss for the combined model is a weighted sum of the reconstruction loss and the Wasserstein loss. The combined loss can be written as follows:
Lcombined =λrecon Lrecon (xi,x^i)+λWasserstein LWasserstein (xi,x^i),where λrecon and λWasserstein are hyperparameters controlling the relative importance of the reconstruction loss and the Wasserstein loss. The generator is trained to minimize this combined loss. The workflow of WGAN-AE is summarized in Algorithm 1.
E. Training Procedure
The training procedure alternates between updating the discriminator and the generator:
Discriminator Update: The discriminator is trained to distinguish between real and fake data by minimizing Ldisc . The real data samples xi are drawn from the dataset, and the fake data samples x^i are generated by the autoencoder.
Generator Update: The generator is updated to minimize the combined loss Lcombined, which includes both the reconstruction error and the Wasserstein loss. This is done by training the generator to fool the discriminator and simultaneously reconstruct the input data accurately.
The discriminator’s parameters ψ are updated using a standard optimization algorithm, while the generator’s parameters θ and ϕ are updated jointly. The training process is summarized in the Algorithm 2.
Experiments and Results
The proposed WGAN-AE architecture is implemented and evaluated in the Google Colab Pro platform. To ensure the optimal training of WGAN-AE, we selected the range of suitable hyperparameters through the hit and trial method. The customized and default hyperparameters utilized in training are presented in Tables 2 and 3, respectively. The following provides a brief discussion of the experimental procedures and an in-depth analysis of experimental outcomes.
Customized hyperparameters for the training of proposed WGAN-AE
Parameter
Value
Encoding dimension
14
Discriminator layers
128, 64 neurons (ReLU activation)
Generator activation
ReLU (encoding), Sigmoid (decoding)
Loss function (GAN)
Wasserstein Loss
Loss function (Final Model)
Mean squared error, Categorical Crossentropy (loss weights: 0.5, 0.5)
Optimizer
Adam
Batch size
256
Epochs (GAN Training)
10
Epochs (K-Fold Training)
5
K-Fold splits
5
Random seed
42
Default hyperparameters for the training of proposed WGAN-AE
Parameter
Default value
Learning rate (Adam Optimizer)
0.001
Beta1 (Adam Optimizer)
0.9
Beta2 (Adam Optimizer)
0.999
Epsilon (Adam Optimizer)
1e-7
Weight initialization (Dense Layers)
Xavier initialization
Performance Evaluation with 5G-NIDD Dataset
The 5G-NIDD dataset is an important benchmark for evaluating the performance of intrusion detection systems in next-generation 5G-enabled IoT networks, characterized by speed, low latency, and a variety of threats. The following presents a brief discussion of the experimental results.
Cross-Validation Performance
To ensure robustness, the model was subjected to five-fold cross validation to prevent overfitting on a single dataset. The five-fold cross-validation results are shown in Fig. 4. The experimental outcomes delivered quite impressive Precision-Recall AUC scores ranging from 99.79% to 99.91%, with an average of 99.87%. This is because the model is very precise, with low false positive and false negative rates for high accuracy. The scores for the Balanced Accuracy ranged from 98.79% to 98.97% with an average of 98.91%. These cross validation results show that the model is highly generalizable with stable performance across different data splits.
5-fold cross validation with 5G-NIDD dataset
Test Set Performance
To further assess the model’s effectiveness, it was evaluated on an independent test set, where it achieved a notable performance score. The accuracy was 97.35%, with precision, recall and F1-score all being very close to each other at 97.39%, 97.35%, and 97.35, respectively. These almost equal values across the different metrics indicate that the model has a good center that helps it avoid false positives and negatives. The high F1 score further confirms that both precision and recall are good, so the model does not give many false positives while also detecting malicious traffic effectively.
Multiclass Performance Analysis
The confusion matrix presented in Fig. 5 provides deeper insights into the model’s classification performance. It shows that the model made few misclassifications, particularly for well-defined attack patterns such as UDP Flood, HTTP Flood, and SYN Scan. However, some minor misclassifications were observed between attacks with similar characteristics, such as SYN Flood and TCP Connect Scan, where the model occasionally confused connection-based attacks due to their similar network behavior profiles. Despite these minor misclassifications, the overall error rate was low, reinforcing the model’s high reliability. Fig. 6 presents a detailed multiclass performance evaluation.
Confusion matrix for 5G-NIDD dataset
Multiclass performance evaluation with 5G-NIDD dataset
Computational Performance Analysis
The computational performance analysis of the proposed WGAN-AE model on the 5G-NIDD dataset demonstrates efficient processing capabilities. The training phase was completed in 59.20913 s, indicating that the model can handle complex data with moderate computational requirements. During the inference stage, the model processed the entire test set in 81.51074 s, reflecting its capacity to analyze and generate predictions efficiently. The per-sample latency was measured at 0.06704 ms, ensuring minimal delay during inference. Moreover, the throughput of the model was recorded at 14,916.93066 samples per second, highlighting its capability to process a large volume of data in heterogeneous IoT networks.
Attack Detection Time Analysis
In real-world applications, fast detection of attacks is crucial to mitigate threats in real time. The proposed model demonstrated impressive low-latency performance across various attack types, processing most attacks within 0.06 to 0.08 milliseconds. This ensures that it can quickly identify and respond to malicious activities without causing significant delays. However, the ICMP Flood attack took slightly longer to detect at 0.29605 ms, likely due to its bursty nature and the larger packet sizes involved, which required additional computational resources. Despite this, all other attack detection times remained within sub-millisecond latencies, confirming that the model is well-suited for real-time intrusion detection in 5G-enabled IoT networks. Fig. 7 illustrates each class’s attack detection time.
Detection time of each individual class in 5G-NIDD dataset
Model Size and Deployment Feasibility
Considering the limited storage and computational resources available on many IoT devices, the model’s memory footprint is a key factor for deployment feasibility. The proposed model has a compact size of just 60.24 kB, making it highly efficient for deployment on resource-constrained IoT devices. The smaller memory footprint, high detection accuracy, and low latency make it an ideal choice for deployment in real-world IoT networks.
Performance Evaluation with IDSIoT2024 Dataset
The IDSIoT2024 dataset contains diverse IoT traffic data to evaluate IDSs in dynamic and resource-constrained IoT networks. The following provides a detailed analysis of experimental results with the IDS IoT 2024 dataset.
Cross-Validation Performance
The five-fold cross-validation results are presented in Fig. 8. The Precision-Recall AUC scores for the five folds were very high, with an average of 94.09%, ranging from 91.81% to 95.66%. This shows that the model is capable of a good precision-recall tradeoff, i.e., it can avoid many false positives and false negatives. The Balanced Accuracy values were also very good, with a range of 88.54%–91.59% and an average of 90.53%. These results show that the model performs reliably in distinguishing attack and benign traffic without bias and thus is likely to be deployable in a generalizable manner across IoT networks.
5-fold cross validation with IDSIoT2024 dataset
Test Set Performance
After training the model using the training dataset and validating its efficiency using the validation set, it performed exceptionally on an independent test set with high accuracy of 98.38% to distinguish between normal and malicious traffic. All other metrics like precision, recall and F1 scores were also very close to each other, with values of 98.34%, 98.38% and 98.27%, respectively. This indicates that the model performs well without being too sensitive or ignoring real anomalies. Furthermore, the decision-making process of the model is straightforward, making it easy to interpret and trust the results.
Multiclass Performance Analysis
The confusion matrix in Fig. 9 gives a more accurate view of the model’s classification accuracy. When the multi-class evaluation was performed, the model’s performance was found to be consistent across all the other classes except the ‘Injection’ class. This means that although the model is very good at identifying different attacks, there could be some difficulties in distinguishing between some of the attacks, especially Injection attacks, which may have attack patterns that are similar to those of other malicious activities. The low performance in this case can be ascribed to the characteristics of Injection attacks which are often quiet and their traffic is not easily distinguishable. Nonetheless, the general multiclass performance is good which indicates that the model is well positioned to deal with different kinds of attacks. The multiclass performance is detailed in Fig. 10.
Confusion matrix for IDSIoT2024 dataset
Multiclass performance evaluation with IDSIoT2024 dataset
Computational Performance Analysis
For the IDSIoT2024 dataset, the proposed WGAN-AE model exhibited a significantly lower training time of 10.67866 s, suggesting that the model adapts efficiently to this dataset. The inference time for the complete test set was remarkably fast at 6.43554 s, emphasizing the model’s ability to handle large-scale IoT data effectively. The latency per sample was recorded at 0.06684 ms, demonstrating minimal delay during prediction. Additionally, the throughput achieved was 14,960.67771 samples per second, reflecting a high data processing rate, which is crucial for real-time IoT applications. These results affirm that the WGAN-AE model is computationally efficient across diverse datasets, making it well-suited for high-throughput environments.
Attack Detection Time Analysis
The model demonstrated impressive efficiency with attack detection times, processing most attacks in sub-millisecond times, ranging from 0.05841 to 0.11560 ms. The shortest detection times were observed for normal traffic and attacks like DoS and Routing, indicating the model’s ability to detect these traffic types quickly. The slightly longer detection times for more complex attacks, such as Injection (0.10729 ms) and MITM (0.11560 ms), still remained well within acceptable thresholds for real-time monitoring. These results highlight the model’s suitability for use in environments where prompt threat detection is essential. Fig. 11 illustrates each class’s attack detection time.
Detection time of each individual class in IDSIoT2024 dataset
Model Size and Deployment Feasibility
The model’s compact size of 61.84 kB makes it highly suitable for deployment in resource-constrained IoT environments. This small memory footprint ensures that the model can be easily integrated into IoT devices with limited computational resources. The model’s lightweight nature, high detection accuracy, and low detection latency demonstrate its potential for deployment in real-world IoT networks, where both efficiency and security are critical. Given the increasing demand for effective and resource-efficient intrusion detection in IoT systems, this model offers a promising solution for ensuring the security of IoT devices.
Performance Comparison of Proposed WGAN-AE with State-of-the-Art GAN Variants
To analyze the efficacy of the proposed WGAN-AE, we compared the performance with state-of-the-art GAN variants, including vanilla GAN, conditional GAN, least square GAN, Information Maximizing Generative Adversarial Networks (info GAN), and boundary equilibrium GAN. To ensure a fair comparison, we implemented all these GAN models on similar experimentation platforms with similar datasets. The following presents a detailed comparative analysis.
Detection Rate and False Alarm Rate
The WGAN-AE consistently outperforms other GAN variants in detection accuracy while maintaining lower false alarm rates across different attack classes. Table 4 presents comparative evaluation results for the 5G-NIDD dataset. The proposed scheme achieved a remarkable 100% detection rate for ICMPFlood attacks with zero false alarms, a stark contrast to Vanilla GAN’s 29.69% detection rate. Similarly, for SYNFlood and SYNScan attacks, WGAN-AE achieves near-perfect detection, significantly surpassing alternatives such as Least Squares GAN and Boundary Equilibrium GAN. The improvement is also evident in benign traffic classification, where WGAN-AE attains a detection rate of 95.84% with a reduced false alarm rate of 1.93%, demonstrating its robustness in distinguishing normal and attack traffic.
Comparative analysis of detection rate (DR) vs. false alarm rate (FAR) for 5GNIDD Dataset (all values are in percentage, %)
Class
Vanilla GAN
Conditional GAN
Least Squares GAN
Info GAN
Boundary Equilibrium GAN
WGAN-AE (Proposed Scheme)
DR
FAR
DR
FAR
DR
FAR
DR
FAR
DR
FAR
DR
FAR
Benign
87.094
6.529
82.563
8.047
79.527
6.197
93.227
3.683
87.942
4.784
95.847
1.927
HTTPFlood
97.532
0.635
98.203
1.074
97.016
0.753
98.455
0.338
97.566
0.517
99.471
0.135
ICMPFlood
29.697
0.002
0.000
0.000
0.087
0.000
98.788
0.002
61.385
0.005
100.00
0.000
SYNFlood
85.876
0.019
82.636
0.028
83.870
0.025
86.092
0.010
85.639
0.011
99.928
0.003
SYNScan
95.375
0.019
92.616
0.050
92.840
0.043
99.606
0.008
97.296
0.016
99.776
0.006
SlowrateDoS
90.687
0.331
84.521
0.239
88.885
0.375
94.907
0.181
92.338
0.300
98.009
0.065
TCPConnectScan
98.499
0.188
96.225
0.252
96.848
0.245
98.998
0.116
98.768
0.160
99.771
0.002
UDPFlood
89.609
8.205
87.211
11.147
90.233
13.147
94.137
4.249
92.384
7.627
96.912
2.604
UDPScan
93.015
0.019
87.583
0.044
85.924
0.019
99.453
0.005
96.259
0.020
99.642
0.002
Table 5 presents the comparative analysis for the IDSIoT2024 dataset. A similar trend is observed in the IDSIoT2024 dataset, where WGAN-AE exhibited high detection rates for Routing (99.35%) and DoS (99.05%) attacks, outperforming all other models. The false alarm rate remains consistently low, reinforcing its reliability for real-world deployment. However, a noticeable weakness emerges in detecting Injection attacks, where it underperformed compared to Information Maximising Generative Adversarial Networks (InfoGAN).
Comparative analysis of detection rate (DR) vs. false alarm rate (FAR) for IDSIoT2024 Dataset (all values are in percentage, %)
Class
Vanilla GAN
Conditional GAN
Least squares GAN
Info GAN
Boundary equilibrium GAN
WGAN-AE (Proposed Scheme)
DR
FAR
DR
FAR
DR
FAR
DR
FAR
DR
FAR
DR
FAR
DoS
94.093
8.451
97.008
4.893
96.200
14.357
97.993
1.923
95.230
6.198
99.053
0.565
Injection
34.063
0.078
43.125
0.073
35.859
0.061
52.656
0.118
36.484
0.058
40.156
0.093
MITM
100.000
0.340
100.00
0.340
100.00
0.340
100.000
0.340
100.00
0.340
100.00
0.340
Malware
99.180
0.404
99.410
0.210
99.380
0.358
99.190
0.072
99.380
0.298
99.420
0.294
Normal
82.980
0.935
88.520
0.866
83.510
1.159
91.980
0.935
84.880
0.794
91.030
0.688
Routing
65.620
2.174
78.330
0.811
33.280
0.806
92.090
0.422
77.100
1.781
99.350
0.010
Vuln_Analysis
95.510
1.514
97.610
1.332
94.825
1.713
97.590
1.085
95.920
1.401
98.365
1.205
Computational Efficiency and Resource Utilization
One of the promising features of WGAN-AE is its efficient training process, which is significantly faster than traditional GANs. The computational efficiency and resource utilization comparison are presented in Tables 6 and 7 for 5G-NIDD and IDSIoT2024 datasets, respectively. On the 5G-NIDD dataset, the training time is reduced to 59.2 s, compared to InfoGAN’s 214.9 s and Vanilla GAN’s 208.2 s. This efficiency stems from incorporating an autoencoder, which compresses input data before training, substantially reducing computational complexity. A similar advantage is observed in the IDSIoT2024 dataset, where the training time is just 10.67 s, making WGAN-AE faster training models among its peers. Another notable advantage of WGAN-AE is its remarkably small model size. The model requires only 0.060 MB for 5G-NIDD and 0.061 MB for IDSIoT2024, making it an excellent choice for deployment in resource-constrained IoT environments.
Comparative analysis of training cost, inferencing time, latency, throughput, and model size for 5GNIDD dataset
GAN variants
Training cost (sec)
Inferencing time (sec)
Latency (ms)
Throughput (Samples/sec)
Model size (MBs)
Vanilla GAN
208.23392
61.12484
0.05027
19891
0.78463
Conditional GAN
112.46435
60.25724
0.04956
20178
0.78463
Least squares GAN
155.00871
60.92937
0.05011
19955
8.65162
Info GAN
214.93845
61.21826
0.05035
19861
0.78463
Boundary equilibrium GAN
208.44314
71.86630
0.05911
16918
0.78463
WGAN-AE
59.20913
81.51074
0.06704
14916
0.06024
Comparative analysis of training cost, inferencing time, latency, throughput, and model size for IDSIoT2024 dataset
GAN variants
Training cost (sec)
Inferencing time (sec)
Latency (ms)
Throughput (Samples/sec)
Model size (MBs)
Vanilla GAN
27.15654
5.66325
0.05882
17000
0.79781
Conditional GAN
19.73150
5.18062
0.05381
18584
0.79781
Least squares GAN
32.70655
5.26421
0.05468
18289
8.70435
Info GAN
26.47308
4.95029
0.05142
19449
0.79781
Boundary equilibrium GAN
24.77814
5.02102
0.05215
19175
0.79781
WGAN-AE
10.67866
6.43554
0.06684
14960
0.06184
Trade-offs and Limitations: A Balanced Perspective on WGAN-AE’s Promising Performance
The WGAN-AE model demonstrates remarkable performance in terms of detection accuracy and false alarm reduction, significantly outperforming state-of-the-art GAN variants across diverse attack classes. While a detailed evaluation reveals some trade-offs in inferencing time and throughput, these differences are relatively minor compared to the notable improvements WGAN-AE brings in detection performance.
Latency and Inferencing Time: Minimal Trade-Offs for Superior Detection
A closer analysis of the inferencing time and latency highlights that although WGAN-AE incurs a slightly higher latency (0.067 ms for 5G-NIDD and 0.066 ms for IDSIoT2024) compared to InfoGAN and Vanilla GAN (around 0.051 ms), the difference is negligibly small in practical scenarios. This minimal latency overhead is a small price for achieving significantly higher detection rates and lower false alarm rates. In real-world intrusion detection environments, where accurate and reliable attack classification is paramount, this marginal increase in latency does not compromise the system’s overall responsiveness.
Throughput: Prioritizing Accuracy over Speed in High-Stakes Scenarios
Similarly, while WGAN-AE demonstrates slightly lower throughput (around 14,916 samples/sec for 5G-NIDD and 14,960 samples/sec for IDSIoT 2024) compared to InfoGAN and Conditional GAN (which process over 19,000 samples/sec), this trade-off is more than compensated for by the model’s superior detection performance. Although important in high-speed environments, it becomes secondary in scenarios where accuracy and reliability are critical. For instance, in mission-critical IoT or 5G networks, ensuring that malicious traffic is identified with near-zero false alarms is more desirable than marginally higher processing speed.
Detection Superiority and False Alarm Reduction: WGAN-AE’s Competitive Edge
The strength of WGAN-AE lies in its ability to consistently achieve higher detection rates across a wide range of attack types, including difficult-to-detect threats such as ICMPFlood, SYNFlood, TCPConnectScan, and UDPScan, while simultaneously maintaining a significantly lower false alarm rate. For example, on the 5G-NIDD dataset, WGAN-AE achieves a 100% detection rate for ICMPFlood attacks with a 0% false alarm rate, outperforming all other models. Even for complex attack types in the IDSIoT2024 dataset, WGAN-AE maintains exceptional performance, highlighting its robustness and reliability in detecting both known and emerging threats.
Future Directions: Enhancing WGAN-AE for Greater Efficiency
While WGAN-AE has already set a high standard in intrusion detection, a few strategic enhancements could further refine its performance regarding inferencing time and throughput. The following two recommendations can help address these minor trade-offs:
Model Pruning and Lightweight Architectures for Faster Inferencing
Model pruning and lightweight architectures can be employed to minimize inferencing time and latency without sacrificing detection accuracy. Pruning reduces model complexity by removing redundant connections and neurons, leading to a lighter and faster network while retaining essential feature representations. Additionally, incorporating quantization techniques can further reduce model size and computation requirements, making WGAN-AE more efficient for real-time applications. This approach can maintain the model’s detection superiority while ensuring faster inferencing in large-scale or latency-sensitive environments.
Parallel Processing and Distributed Inference for Higher Throughput
Implementing parallel processing techniques and distributed inference frameworks can significantly improve performance in high-speed networks, enhancing throughput and enabling real-time intrusion detection. WGAN-AE can process a larger volume of samples concurrently by partitioning incoming network traffic across multiple computational nodes, thereby increasing overall throughput. Additionally, leveraging edge-cloud hybrid architectures can offload preliminary anomaly detection to edge devices, reducing the computational burden on central servers while maintaining accuracy and reliability.
Conclusion
This paper proposed a hybrid framework for an efficient IDS for IoT networks using GAN and autoencoder architectures. The proposed WGAN-AE successfully identified a range of cyberattacks with higher accuracy. The performance of the designed IDS framework was evaluated using two open source datasets, 5GNIDD and IDSIoT2024. The experimental outcomes confirm the higher attack detection accuracy in both 5-fold cross-validation scenarios and with respect to independent testing data. The microseconds attack detection time for each class and the low memory footprint makes it suitable for deployment in resource constrained IoT devices and networks.
The authors extend their appreciation to the Deanship of Research and Graduate Studies at King Khalid University for funding this work through Large Group Project under grant number (RGP.2/245/46). This work is funded by Princess Nourah bint Abdulrahman University Researchers Supporting Project number (PNURSP2025R760), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia. The research team thanks the Deanship of Graduate Studies and Scientific Research at Najran University for supporting the research project through the Nama’a program, with the project code NU/GP/SERC/13/352-1.
Funding Statement
The authors extend their appreciation to the Deanship of Research and Graduate Studies at King Khalid University for funding this work through Large Group Project under grant number (RGP.2/245/46). This work is funded by Princess Nourah bint Abdulrahman University Researchers Supporting Project number (PNURSP2025R760), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia. The research team thanks the Deanship of Graduate Studies and Scientific Research at Najran University for supporting the research project through the Nama’a program, with the project code NU/GP/SERC/13/352-1.
Author Contributions
Mohammed S. Alshehri: Writing an original draft, Visualization, Validation, Software, Project administration, Methodology, Investigation, Formal analysis, Conceptualization. Oumaima Saidani: Writing an original draft, Visualization, Methodology, Investigation, Formal analysis, and Conceptualization. Wajdan Al Malwi: Writing, review & editing, Writing an original draft, Visualization, Validation. Fatima Asiri: Writing, review & editing, Visualization, Validation. Shahid Latif: Writing an original draft, Software, Methodology, Formal analysis. Aizaz Ahmad Khattak: Review & editing, Visualization, Methodology. Jawad Ahmad: Review & editing, Software, Project administration. All authors reviewed the results and approved the final version of the manuscript.
Availability of Data and Materials
We are happy to share the processed datasets and Jupyter Notebooks of the proposed scheme for research purposes upon request, subject to the approval of our research group.
Ethics Approval
Not applicable.
Conflicts of Interest
The authors declare no conflicts of interest to report regarding the present study.
ReferencesMehediST, AnwarA, RahmanZ, AhmedK, IslamR. Dependable intrusion detection system for IoT: a deep transfer learning based approach. IEEE Trans Indus Inform. 2022;19(1):1006–17. doi:10.1109/TII.2022.3164770.ShafiqM, GuZ, CheikhrouhouO, AlhakamiW, HamamH. The rise of “Internet of Things”: review and open research issues related to detection and prevention of IoT-based security attacks. Wirel Commun Mob Comput. 2022;2022(1):8669348. doi:10.1155/2022/8669348.JarwarMA, FREngJWC, AliS. Modelling industrial IoT security using ontologies: a systematic review. IEEE Open J Commun Soc. 2025;6(3):2792–821. doi:10.1109/OJCOMS.2025.3532224.LatifS, HumaZ, JamalSS, AhmedF, AhmadJ, ZahidA, et al. Intrusion detection framework for the internet of things using a dense random neural network. IEEE Trans Indus Inform. 2021;18(9):6435–44. doi:10.1109/TII.2021.3130248.AlwaisiZ, SoderiS. Towards robust IoT defense: comparative statistics of attack detection in resource-constrained scenarios. In: EAI International Conference on Body Area Networks; 2024 Feb 4–5; Milan, Italy. p. 272–91.RahmanS, PalS, MittalS, ChawlaT, KarmakarC. SYN-GAN: a robust intrusion detection system using GAN-based synthetic data for IoT security. Internet of Things. 2024;26(7):101212. doi:10.1016/j.iot.2024.101212.LinR, QiuH, WangJ, ZhangZ, WuL, ShuF. Physical layer security enhancement in energy harvesting-based cognitive internet of things: a GAN-powered deep reinforcement learning approach. IEEE Internet of Things J. 2024;11(3):4899–913. doi:10.1109/JIOT.2023.3300770.KarthikaS, DurgadeviM. Generative Adversarial Network (GAN): a general review on different variants of GAN and applications. In: 2021 6th International Conference on Communication and Electronics Systems (ICCES). Coimbatre, India: IEEE; 2021. p. 1–8.FetayaE, JacobsenJH, GrathwohlW, ZemelR. Understanding the limitations of conditional generative models. arXiv:190601171. 2019.Cabezon PedrosoT, SerJD, Díaz-RodríguezN. Capabilities, limitations and challenges of style transfer with CycleGANs: a study on automatic ring design generation. In: International Cross-Domain Conference for Machine Learning and Knowledge Extraction; 2022 Aug 23–26; Vienna, Austria: Springer. p. 168–87.HasanMN, JanSU, KooI. Wasserstein GAN-based digital twin-inspired model for early drift fault detection in wireless sensor networks. IEEE Sens J. 2023;23(12):13327–39. doi:10.1109/JSEN.2023.3272908.CaiZ, DuH, WangH, ZhangJ, SiY, LiP. One-dimensional convolutional wasserstein generative adversarial network based intrusion detection method for industrial control systems. Electronics. 2023;12(22):4653. doi:10.3390/electronics12224653.AlrayesFS, ZakariahM, AminSU, KhanZI, HelalM. Intrusion detection in IoT systems using denoising autoencoder. IEEE Access. 2024;12:122401–25.BoppanaTK, BagadeP. GAN-AE: an unsupervised intrusion detection system for MQTT networks. Eng Appl Artif Intell. 2023;119(11):105805. doi:10.1016/j.engappai.2022.105805.LiS, CaoY, LiuS, LaiY, ZhuY, AhmadN. HDA-IDS: a hybrid DoS attacks intrusion detection system for IoT by using semi-supervised CL-GAN. Expert Syst Appl. 2024;238(15):122198. doi:10.1016/j.eswa.2023.122198.de Araujo-FilhoPF, KaddoumG, CampeloDR, SantosAG, MacêdoD, ZanchettinC. Intrusion detection for cyber-physical systems using generative adversarial networks in fog environment. IEEE Internet of Things J. 2020;8(8):6247–56. doi:10.1109/JIOT.2020.3024800.ZeghidaH, BoulaicheM, ChikhR, BamhdiAM, BarrosALB, ZeghidaD, et al. Enhancing IoT cyber attacks intrusion detection through GAN-based data augmentation and hybrid deep learning models for MQTT network protocol cyber attacks. Cluster Comput. 2025;28(1):58. doi:10.1007/s10586-024-04752-5.DasS, MajumderA, NamasudraS, SinghA. Intrusion detection using CTGAN and lightweight neural network for Internet of Things. Expert Syst. 2025;42(2):e13793. doi:10.1111/exsy.13793.WangH, KandahF, MendisT, MeduryL. Clustering-based intrusion detection system meets multi-critics generative adversarial networks. IEEE Internet Things J. 2025. doi:10.1109/JIOT.2025.3533918.DongB, WangH, LuoR. MasqueradeGAN-GP: a generative adversarial network framework for evading black-box intrusion detection systems. Internet Technol Lett. 2025;16(8):e640. doi:10.1002/itl2.640.BrabinDD, KumarKK, SunithaT. Strengthening security in IoT-based smart cities utilizing cycle-consistent generative adversarial networks for attack detection and secure data transmission. Peer Peer Netw Appl. 2025;18(2):79. doi:10.1007/s12083-024-01838-0.SamarakoonS, SiriwardhanaY, PorambageP, LiyanageM, ChangSY, KimJ, et al.5G-NIDD: a comprehensive network intrusion detection dataset generated over 5G wireless network. IEEE Dataport. 2022. doi:10.21227/xtep-hv36.ManasaK, Leo JosephLMI. A real time dataset “IDSIoT2024” for machine learning/deep learning based cyber attack detection system for IoT architecture. In: 2025 3rd International Conference on Intelligent Data Communication Technologies and Internet of Things (IDCIoT). Bengaluru, India: IEEE; 2025. doi:10.21227/gfaz-t124.ArjovskyM, ChintalaS, BottouL. Wasserstein generative adversarial networks. In: International Conference on Machine Learning. Sydney, Australia: PMLR; 2017. p. 214–23.