Recognition of otolith shape provides crucial economic and ecological insights into the status of fish populations worldwide. In fact, otoliths serve as an effective tool for promoting environmental sustainability. By analyzing otoliths, researchers can determine fish species stocks, estimate fish age, and conduct taxonomy studies. Recently, the most effective models for otolith image classification have been based on deep neural networks. However, the main challenge with these approaches is their generalizability - the capacity of well-trained models to accurately predict novel data. This raises the question: how can models maintain high performance and efficiency across different datasets? In this paper, we investigate the use of adversarial training methods and the evolution of decision boundaries in deep learning models. Specifically, we investigate how the decision boundary changes throughout the training process of a deep neural network. We build a convolutional neural network (CNN) model to classify otolith images. In addition to the original images, we generate a perturbed version of each image, maintaining the same label. Our results demonstrate that the CNN model can classify 15 types of otoliths with high accuracy. We show that, in adversarial training, the distance between the images and the decision boundary is greater compared to standard training. Moreover, our approach proves to be efficient, achieving an accuracy rate of 92 %.