Ear biometrics has been gaining attention as a reliable method for identity verification, because of its stability over time, non-intrusive nature, and resilience to changes in facial expressions, obstructions. In this study, we explore the use of different deep learning algorithms and compare them for ear recognition that uses generative data augmentation to overcome the challenges of limited datasets and intra-class variability. We employ two advanced generative adversarial network (GAN) models—Conditional Deep Convolutional GAN (C-DCGAN) and Conditional Wasserstein GAN with Gradient Penalty (C-WGAN-GP) to generate realistic synthetic ear images. These images enhance the diversity of the dataset. By combining synthetic data as well as traditional augmentation methods, we train three leading convolutional neural networks: VGG-19, ResNet-152, and EfficientNet-V2-M for classification purposes. Our experiments, conducted on the IIT Delhi and UERC-2019 benchmark datasets, show that EfficientNet-V2-M, when trained with data augmented by WGAN-GP, achieves the highest classification accuracies of 98.6% and 95.3%, respectively. The results show that using GAN-based data augmentation really helps improve how well the classification works, particularly when there isn’t much training data available. When properly fine-tuned, this method offers a practical and reliable way to handle real-world ear biometric systems that can easily scale up as needed.