As the reliance on online payment systems grows, so does the urgency to combat fraudulent activities. In this study, we will analyze and contrast between three different architectures of the models which are designed to prevent online fraud in payments. The models use modernly combined data preprocessing techniques and deep learning architectures, which provide solutions to needs of the imbalanced data sets. The first method incorporates ADASYN algorithm with CNNs to discriminate the transactions from this category. The second method is based on the combination of a Synthetic Minority Over-sampling Technique (SMOTE) and Convolution neural networks type of neural networks. The next part is quite innovative, and it involves the application of SMOTE with an autoencoder, then the model used to classify the data. Our experiments, conducted on a comprehensive dataset of online transactions, reveal the efficacy of each model in detecting fraudulent activities. We analyse and compare the models’ performance based on metrics such as accuracy, precision, recall, and F1 score and observed that our third approach (novel) outperformed the first two models.