Computer vision applications require substantial amounts of data for effective training and accurate inference in a variety of tasks. However, in many real-world scenarios, data insufficiency is a common issue, which can arise due to various factors such as the rarity of certain conditions, difficulties in data collection, or high costs associated with data acquisition. This insufficiency often leads to computational models with inadequate performance, particularly in terms of their generalization capabilities. Traditional data augmentation techniques are widely used to mitigate overfitting and improve model robustness by artificially increasing the diversity of the training data. However, the application of these techniques is not always feasible or desirable, especially in cases where the augmented data does not accurately represent the underlying distribution. In response to these challenges, this paper explores an alternative data augmentation approach specifically designed for classification tasks. The method leverages adversarial images generated using the Fast Gradient Sign Method (FGSM) with added noise to address sample imbalance and enhance classifier performance. The technique was validated on a dataset of images related to the classification of diseases in coffee plants caused by nutrient deficiencies in the soil. The experimental results demonstrated a significant improvement in model performance, highlighting the effectiveness of the proposed method as a viable alternative to traditional data augmentation techniques.