Original title: Prediction of bank transaction fraud using TabNet an adaptive deep learning architecture
Prashanth BS, Manoj Kumar, Ariful Hoque, Nasser Al Muraqab +3 · July 2026
Online banking has led to an increase in fraudulent operations, creating a major problem for banks. Researchers wanted to build an accurate, transparent system to catch these crimes while helping banks follow regulations. They tested five deep learning architectures on a dataset of real bank transactions from India. The models included DNN, GRU, LSTM, CNN1D, and TabNetA deep learning model designed for tables of numbers, which uses attention to pick the most important features for each decision.. To make the training fair, they used an oversampling technique called SMOTEA technique that creates synthetic examples of the minority class so the data isn't so imbalanced. to balance the rare fraud examples with regular transactions. TabNet works by using sparse feature selection. This design choice helps the model focus on the most important pieces of data, making its decisions easier to interpret. In tests, TabNet clearly outperformed the other models. It reached an accuracy of 97.39 % and an ROC-AUCA score from 0 to 1 that measures how well a model separates the two classes; 1 is perfect, 0.5 is random guessing. score of 0.9739. It also created fewer false positives and negatives, making it practical for real-time banking systems.
Imagine sorting a huge box of mixed LEGO bricks by color and shape before starting to build. If you sort everything perfectly, you waste hours and still might lose the one rare piece; if you skip sorting, you dig through the box every time you need something. Here, the bank's transactions are the mixed bricks, and TabNet is the sorter that focuses only on the few details that signal fraud, ignoring the rest.
Banks face rising online fraud, and every missed fraud costs money while every false alarm annoys customers. A model that is both accurate and explainable lets banks catch suspicious transactions and show regulators why a decision was made. That builds trust and supports compliance.
The abstract doesn't say. It reports results on a single Kaggle dataset of Indian bank transactions, so generalizability to other countries, time periods, or live data is unknown. It also doesn't compare against non-deep-learning baselines, and it doesn't report training cost, latency, or real-world deployment results.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Prediction of bank transaction fraud using TabNet an adaptive deep learning architecture — Prashanth BS, Manoj Kumar, Ariful Hoque et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.