Generative Adversarial Networks (GANs) in Networking: A Comprehensive Survey & Evaluation

1. Introduction

• Network growth and 5G demands challenge conventional methods, while scarce training data, class imbalance, and costly labeling complicate network learning.Generative approaches are considered for creating data with real-data properties to improve learning performance.Source: §1 · p. 1

1.1. Previous Work

• Existing GAN surveys mainly cover computer vision, motivating a detailed survey of applications in computer and communication networks.Some surveys touch on other fields, but one cited review covers them only briefly.Source: §1.1 · p. 2, §1 · p. 2

1.2. Key Contributions

• The paper compares GAN architectures, organizes applications into five categories, and proposes a non-image evaluation framework tested on four network datasets.It also discusses open challenges and future research directions.Source: §1.2 · p. 3

1.3. Survey Organization

• The paper proceeds from generative-model background and GAN variants to networking applications, model evaluation, and challenges and future work.Source: §1.3 · p. 3

• A generative model estimates a data probability distribution and draws samples, either by explicitly specifying the distribution or generating samples without defining it.Source: §2 · p. 3
• Generative models are used for filling in missing data and imputation, and can support semi-supervised learning when only some training examples are labeled.Source: §2 · p. 4
• GANs are implicit-density models that define a stochastic sampling process rather than a tractable likelihood, and generate samples in a single step.Explicit-density models instead specify the data distribution; examples include tractable-density flow models and approximating methods such as variational autoencoders.Source: §2 · p. 4, Figure 1 · p. 4

• GANs aim to produce synthetic data resembling the original through two networks, a Generator and a Discriminator, trained in an adversarial game.Source: §2.1 · p. 4
• The Generator creates counterfeits to deceive the Discriminator, which attempts to recognize them.Source: §2.1 · p. 4
• The Generator and Discriminator learn from each other, and training leaves the Generator aiming to mislead the trained Discriminator.Source: §2.1 · p. 4

2.2. GAN Variations and Architectures

• Vanilla GAN maps latent noise to data using generator and discriminator networks, while the reviewed architectures vary in their components and training approaches.The figure identifies G as common across the models and labels D, E, F, and Q as discriminator, encoder, critic, and Q-network.Source: §2.2 · p. 4, Figure 2 · p. 5

2.2.1. Conditional GAN

• Conditional GAN feeds encoded class labels alongside prior noise to the generator and discriminator, enabling class-conditioned sample generation.Its cost function is described as similar to Vanilla GAN’s.Source: §2.2.1 · p. 6

2.2.2. Auxiliary Classifier GAN

• ACGAN supplies class labels and noise to the generator, while its discriminator also predicts labels through a multi-task architecture.Unlike CGAN, ACGAN predicts class labels rather than conditioning both networks on them.Source: §2.2.2 · p. 6

2.2.3. Wasserstein GAN

• WGAN uses a novel cost function to address Vanilla GAN instability; adding noise to generated samples is also proposed to stabilize training before discriminator input.The cost function is described using a supremum over 1-Lipschitz functions.Source: §2.2.3 · p. 6

2.2.4. WGAN-GP

• WGAN-GP adds a gradient-penalty term to WGAN’s cost function to address the bottleneck of critic weight clipping during training.Source: §2.2.4 · p. 7

2.2.5. InfoGAN

• InfoGAN adds independent conditional factors to generator noise and uses an auxiliary distribution to obtain a variational lower bound for the mutual-information term.Directly maximizing the mutual information is described as difficult because it requires access to the posterior distribution.Source: §2.2.5 · p. 7

2.2.6. Least Square GAN

• Least Square GAN replaces cross-entropy loss with least-squares loss to address the vanishing-gradient problem during learning.Source: §2.2.6 · p. 7

2.2.7. Bidirectional GAN

• BiGAN adds an encoder alongside the generator and discriminator, allowing mapping from real data to the latent noise domain.The standard GAN framework is described as not reversible and unable to map data to a latent representation.Source: §2.2.7 · p. 7

• GAN networking applications are grouped into mobile networks, network analysis, internet of things, physical layer, and cybersecurity, with each category reviewed through its challenges and real-world examples.The paper says the chapter ends with a summary of the reviewed applications.Source: §3 · p. 7

3.1.1. 5G Network Slicing

• GANSlicing combines service-oriented slicing with GAN-based prediction to forecast IoT application resource demands and target improved user QoE.Its global resource view considers physical and virtual cellular-network capabilities to support dynamic allocation.Source: §3.1.1 · p. 8
• GANSlicing accepts 16% more requests with 12% fewer resources than tenant-oriented slicing in the same service request batch.The survey reports improved service acceptance ratio and overall service quality.Source: §3.1.1 · p. 8

3.1.2. Self-Organizing Networks

• SON learning methods face limited, costly-to-gather data and class imbalance, especially for rare events that are difficult to reproduce.These constraints complicate the training of algorithms for network planning, management, and optimization.Source: §3.1.2 · p. 8
• A cell-outage detection approach combines Vanilla GAN-generated minority-class samples with AdaBoost to rebalance data before classification.Cell outage is rare, so it forms only a small fraction of network measurements and can bias traditional classifiers toward the majority class.Source: §3.1.2 · p. 9
• Hughes et al. generated synthetic call-duration and start-hour records and reported a 3.43% accuracy improvement.The generated and real data differed more in variance than in mean, which the survey links to GANs' difficulty generating realistic outliers.Source: §3.1.2 · p. 9

• Across five GAN architectures, Aho et al. found Vanilla GAN and LSGAN impractical for network-data generation, while WGAN and its variants offered significant performance gains.Their approach generated synthetic live traffic intended to improve the robustness of network-analysis learning algorithms.Source: §3.2 · p. 9
• FlowGAN generates samples for low-count traffic classes, and its MLP-classifier evaluation found that addressing class imbalance can increase traffic-classifier performance.Source: §3.2 · p. 9
• Li et al. use WGAN and WGAN-GP to morph censored flows toward permitted flows, evaluating camouflage with area under curve and IND-CA.Their evaluation used more than 10,000 network flows and six traffic features; IND-CA is defined in the cited passage with a privacy indicator.Source: §3.2 · p. 9
• GCN-GAN combines graph convolution and LSTM layers in its generator to predict subsequent snapshots from a weighted dynamic network's historical topology.Source: §3.2 · pp. 9–10
• TranGAN adds a neural-network classifier to the generator-discriminator structure and uses labeled source-network relationships to improve social-tie prediction in a target network with missing labels.The source and target networks come from different domains.Source: §3.2 · p. 10
• A WGAN-augmented attack-data method is limited to continuous network features because it does not perform well on discrete features such as protocol type, flags, and service.Source: §3.2 · p. 10

3.3.1. Wireless Sensor Networks

• A GAN-powered wireless-sensor middleware uses generated data to confuse attackers, while its discriminator distinguishes real data and detects anomalies for further processing.The WSN need not generate fake data, which the passage says can significantly reduce power consumption.Source: §3.3.1 · p. 10

3.3.2. Indoor Localization and path planning

• Nabati et al. used GANs to augment RSSI fingerprint data, achieving identical benchmark accuracy with as little as 10% of the real data.The survey also reports reduced collection costs and time.Source: §3.3.2 · p. 10
• A wavelet-transform feature DCGAN generated more diverse CSI feature maps, and the larger training set improved localization accuracy over existing fingerprint-database construction methods.Source: §3.3.2 · pp. 10–11
• A GAN trained on user trajectories generated paths reported as more than 99.9% reliable, while its path classifier achieved around 99% accuracy.The approach separately handles localization and wayfinding.Source: §3.3.2 · p. 11

3.3.3. Human Activity Recognition

• A semi-supervised GAN augmented CSI activity data using 50% of each activity class and improved overall classification accuracy by 3% [91].The method addressed difficulty collecting adequate labeled training data and scaled down Log Loss.Source: §3.3.3 · p. 11
• CsiGAN extends the semi-supervised discriminator from k + 1 to 2k + 1 probabilistic outputs and adds manifold regularization for CSI-based behavior recognition.The study used leave-one-subject-out validation on the SignFi and FallDefi datasets.Source: §3.3.3 · p. 11

3.4. Physical Layer

• Communications GAN approximates channel response with adversarial learning and channel autoencoders, but its training assumes a differentiable channel function.The approach learns an encoder/decoder layout and can learn directly on unseen physical channels when an approximation is sufficient.Source: §3.4 · p. 12
• Ye et al. use a conditional GAN to model channel distributions for end-to-end communication, examining AWGN and time-varying Rayleigh fading channels.Pilot information can be added to conditioning to estimate channel state information; the authors suggest extension beyond those channel types.Source: §3.4 · p. 12
• Yang et al. train on raw measurements to model channel impulse responses, while Smith and Downey report high-accuracy approximations across four complex channels.Figure 5 shows the different models tested in; Smith and Downey’s study addresses nonlinearities, memory effects, and non-Gaussian statistics.Source: §3.4 · p. 12, Figure 5 · p. 13
• AC-WGAN achieved a classification rate greater than 95% for indoor UAV detection and was reported to detect UAV signals outdoors from several hundred meters away.Source: §3.4 · p. 13

3.4.1. Cognitive Radio

• RCGAN, a semi-supervised DCGAN framework, improved modulation-recognition accuracy even when the signal-to-noise ratio was under 4 dB.Source: §3.4.1 · p. 13
• GAN-generated wireless signals have the potential to improve spoofing success even when a neural-network classifier defends the receiver.Source: §3.4.1 · p. 14
• 10 real samples instead of 500 kept jamming misdetection probability and false alarm rate within 0.19% and 3.14%, respectively, of the original-data results.CGAN-generated synthetic data was used to reduce data-gathering time and cost.Source: §3.4.1 · p. 14
• GAN-PA balances covert rate and detection error in a cooperative cognitive-radio network, with near-optimal performance reported when network status is minimal.Source: §3.4.1 · p. 14

3.5.1. IDSs, Malware Detection, and Security Systems

• IDS applications include GAN-generated evasion attacks, anomaly synthesis from normal HIDS data, and GIDS detection of known and unknown attacks using normal data.The cited work also proposes a GAN-based defense to improve IDS robustness against adversarial attacks.Source: §3.5.1 · p. 14
• LSTM-GAN uses Android opcode sequences for malware detection, while transferred GAN pre-trains its generator with an autoencoder for settings such as zero-day attacks with little data.Source: §3.5.1 · p. 15
• A Vanilla GAN identified rogue RF transmitters with about 99.99% accuracy; Replay Buffer training and two-touch features support Android pattern-lock anomaly detection.Source: §3.5.1 · p. 15

3.5.2. Malware Adaption and Improvement

• With 217 real network flows trained for 400 epochs, LSTM-GAN-based malware adaptation reduced the blocked malware percentage to zero.The method learns benign Facebook chat traffic features so malware can adapt its traffic accordingly.Source: §3.5.2 · p. 15
• With a neural-network substitute detector as the black box, almost all adversarial examples generated by MalGAN bypassed detection.Figure 6 depicts the MalGAN architecture [122]. MalGAN requires an internally built detector, uses only 128 features, and requires multiple malware instances for training.Source: §3.5.2 · p. 15, Figure 6 · p. 15
• GANs were reported successful in exploratory black-box API attacks even with a minimal quantity of training samples.Source: §3.5.2 · p. 15

3.5.3. Other Security Applications

• Among the tested fraud-detection GANs, RGWAN performed best and vanilla GAN worst, with mode collapse cited for vanilla GAN.Other approaches generate synthetic standard transactions or fraudulent data for training and detection.Source: §3.5.3 · pp. 15–16
• PassGAN guessed between 51% and 73% of new unique passwords compared with other tools, including passwords not generated by password tools.The authors say larger training datasets would require more complex networks, more comprehensive training, and more generated passwords, while calling those costs negligible relative to benefits.Source: §3.5.3 · p. 16

4. Evaluation Framework

• The evaluation framework trains five state-of-the-art GANs on four network-related datasets and applies the proposed metrics and visualizations to assess performance.The datasets represent different types and shapes of network data.Source: §4 · p. 16

4.1. Evaluation Metrics

• The framework uses MMD and Wasserstein distance to compare real and generated samples, since most commonly used GAN metrics apply only to images.The paper also visualizes estimated distributions and uses Q-Q plots as a visual guide, not reliable proof, of similarity.Source: §4.1 · pp. 16, 18, §4.2 · p. 19

4.2. Experimental Methodology and Results

• The evaluation trains Vanilla GAN, CGAN, BIGAN, LSGAN, and WGAN on WiFi RSSI, CSI, KDD99, and RadioML data, comparing them with MMD and EMD.KDD99 training uses 18 continuous non-zero features because GANs cannot generate discrete data; model parameters are not tuned.Source: §4.2 · p. 18

4.3. Results

• No tested GAN has a decisive overall advantage: performance depends on data dimensionality, magnitude, morphology, and type, with LSGAN and Vanilla GAN yielding relatively large MMD and EMD on high-dimensional CSI.They show acceptable performance on other data types; the authors suggest architecture changes and hyperparameter tuning could alter the results.Source: §4.3 · p. 19
• The tested GANs cannot capture very fast fluctuations in data distributions, possibly because of intrinsic model bias; deeper models may help, but such complexity has not yet been applied practically in networking.The authors also suggest changing model depth, layer count, and hyperparameter tuning could significantly change the results.Source: §4.3 · pp. 19–20

5. Discussion and Conclusion

• The survey reviews GAN applications across networking and evaluates several models on datasets drawn from the reviewed papers, alongside outlining current challenges and future work.It groups applications into five categories: mobile networks, network analysis, IoT, physical layer, and cybersecurity.Source: §5 · pp. 20–21

5.1. Take-home lessons

• GANs may benefit networking where data are scarce or imbalanced, and in exceptional or adversarial settings such as malware detection.Applications include traffic classification, modulation recognition, self-organizing networks, spectrum sensing, malware detection, and IoT.Source: §5.1 · p. 21
• For non-image data, researchers can adapt image GANs to multilayer perceptrons with normalized data or convert data into images; which approach is better likely depends on the data.GANs can also support semi-supervised learning by training on unlabeled data and using a small labeled portion to train a discriminator.Source: §5.1 · p. 21

5.2. Future Work

• GANs can be adapted to generate continuous multidimensional numerical data, but they cannot yet convincingly generate arbitrary discrete data in arbitrary environments.The proposed BSGAN has not yet been evaluated on discrete networking data.Source: §5.2 · p. 21
• Generating data that is indistinguishable to both humans and machines remains an open challenge despite improvements to GAN convergence reliability, mode collapse, and instability.The survey also identifies GAN use in reinforcement learning for network-related tasks as work still to be done.Source: §5.2 · pp. 21–22