Enhancing Neural Network Performance with Generative Adversarial Networks

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

General Background Deep neural networks achieve remarkable success in computer vision but remain susceptible to overfitting and data scarcity. Specific Background Generative Adversarial Networks offer a framework to produce high-fidelity synthetic images through competitive zero-sum training between generator and discriminator sub-networks. Knowledge Gap However, the direct impact of generative augmentation on network generalization and adversarial robustness under noisy conditions requires rigorous empirical quantification. Aims This study evaluates how integrating generative synthetic samples into the training pipeline enhances neural network classification accuracy and resilience. Results Incorporating augmented images into CIFAR-10 training improved classification accuracy from 89.7% to 94.3%, while simultaneously increasing resilience against FGSM, PGD, and CW adversarial attacks. Novelty A unified evaluation architecture combining Wasserstein gradient penalty formulations with noisy data testing is established. Implications These findings demonstrate that generative synthetic data effectively mitigates data limitations and stabilizes deep learning deployment in unpredictable operational environments. Key Findings Highlights Synthetic data augmentation increases convolutional neural network accuracy on CIFAR-10 to 94.3%. Adversarial training enhances model resilience against FGSM, PGD, and CW attacks. Generative samples mitigate overfitting and restore classification accuracy under noisy data conditions. Keywords : Generative Adversarial Networks, Neural Networks, Data Augmentation, Synthetic Data, Adversarial Robustness