Credit card fraud detection is a critical challenge for financial institutions, with traditional methods often proving inadequate against evolving fraudulent tactics. In this study, we compare the performance of two artificial neural network architectures—Deep Neural Networks (DNNs) and Convolutional Neural Networks (CNNs)—for detecting credit card fraud. Our investigation focuses on their effectiveness in handling highly unbalanced datasets typical of fraud detection tasks. The DNN model, trained using class_weights option, demonstrates adaptability in learning complex patterns from structured data, achieving notable accuracy and performance metrics. In contrast, the CNN model is trained with SMOTE balanced data and excels in capturing spatial hierarchies of features, particularly beneficial for sequence data like credit card transactions. We evaluate their performance using metrics such as accuracy, recall, precision, F1-score, and Matthew`s correlation coefficient (MCC) crucial for assessing models on unbalanced datasets. Experimental results on a real-world dataset reveal that both models demonstrated strong capabilities in fraud detection. The CNN model with SMOTE significantly outperformed the DNN with class weighting in all key metrics, particularly in terms of precision and MCC. The CNN model`s balanced approach using SMOTE effectively addressed the class imbalance issue, resulting in a model that is highly accurate and reliable in detecting fraudulent transactions without generating excessive false positives.