Davaslioglu and Sagduyu [26] introduced a new approach to spectrum augmentation using GANs. This approach effectively handles the inherent difficulties of applying ML classifiers in the cognitive radio domain. Generative adversarial learning is implemented to obtain and generate synthetic data samples, enhancing classifier accuracy via training data augmentation. Additionally, they demonstrated that the inclusion of high-fidelity training data in the retraining process of the classifier leads to improved accuracy in spectrum sensing. This improvement is comparable to the hypothetical scenario, where extra real data are available within the same spectrum environment. Gong et al. [27] proposed a framework for identifying specific emitters utilizing the InfoGAN approach. To enhance the quality of GANs, the proposed framework contains two additional inputs. The proposed framework involves the creation of an artificial RFF vector. This vector is generated using the histograms of bispectral information extracted from received signals. This synthetic RFF vector aims to improve the distinction between individual elements. Additionally, a structured multimodal latent vector is utilized. This latent vector incorporates prior knowledge of fading channel distributions and was designed to align with the characteristics of received signals. The proposed framework demonstrated an assessment score of 87% for the GAN generator. Truong and Yanushkevich [28] proposed an approach for synthesizing radar signals using GANs. Their research explored the application of GANs in the domain of one-dimensional radar signal data generation and augmentation, which has yet to be thoroughly studied. The approach intends to replicate simplified real scenarios in which suspects endeavor to obscure highly reflective objects underneath multiple layers of clothing. Training samples are used to train GANs to generate samples that closely match the training data distribution. This approach has presented encouraging outcomes in synthesizing indiscernible radar signal samples from training samples.

In the same context, Roy et al. [29] provided complete band spectrum simulations and emulation examples using GANs that are sufficiently realistic. In addition, they considered the need for signal synthesis to validate and show that such an approach was feasible, improve the algorithmic approach and quantify and prove its efficiency with current signal sets. Tang et al. [30] addressed the issue of limited labeled data. They looked into the potential of using GANs to effectively generate images to expand original datasets through data augmentation. The findings of this experiment suggest that the GAN-based data augmentation framework used in this study can potentially enhance CNN classification performance, resulting in accuracy improvement ranging from 0.1–6%. Castelli et al. [31] investigated using GANs to generate synthetic telecommunication data about Wi-Fi signal quality. Vanilla GANs and Wasserstein GAN architectures were employed in this study. According to the results of their experiment, both models can generate synthetic data whose distribution matches that of actual data. Additionally, they demonstrated that an ML classifier showed a limited ability to differentiate between actual and generated data, thus providing further evidence of the resilience of the GAN-based model.

Furthermore, He et al. [32] introduced a communication signal enhancement model to address environmental changes. They developed a model utilizing GANs, which includes an encoding–decoding structure based on convolutional layers. This model’s primary objective is to mitigate noise and interference in signals expeditiously. The experimental findings confirm that the proposed model provides greater efficacy in enhancing communication signals. Zhou et al. [33] developed a generic adversarial framework called the wireless signal enhancement GAN for wireless signal enhancement. Ablation studies were employed to demonstrate the value of each part of the objective’s function. Unlike the signal enhancement method, which learns noise distribution and interference characteristics and then removes them from the original signals, this generator learns the characteristics of signals in an adversarial manner. Patel et al. [34] suggested a data augmentation approach involving the utilization of conditional GANs for automatic modulation classification. Their objective was to utilize conditional GANs to produce high-quality labeled data using just a small quantity of initial data. This approach addresses the significant expenses and difficulties faced in acquiring wireless datasets. Through experimentation using an open-source dataset, they demonstrated that the suggested data augmentation approach can significantly enhance the performance of automatic modulation classification models.