Case study

Payment fraud prevention

As part of a multi-layered fraud strategy, an innovative deep learning model is used to detect new fraud patterns in digital payments.

3 AI experts

2 Solutions Architects

2 Domain Experts

1 Project Manager


New paradigms of fraud prevention

Our client, a financial institution, provides enterprise payment solutions within the EU to support banks, insurance companies, merchants, businesses, and public enterprises.

Banks utilize our client’s card payment services and digital payment acceptance systems in order to support their growing purchasing channels (e-commerce, NFC, mobile, contactless). Merchants use our client’s authorization and payment-management services to facilitate transactions over leading international circuits.

Fraud prevention and fraud management are central to these payment solutions, and Nextbit has been a key partner in making these digital payments more secure.


A deep learning model for real-time fraud detection

As part of a multi-layered fraud strategy, Nextbit has introduced a deep learning model to detect new fraud patterns in digital payments. This innovative algorithm is focused on detecting patterns not seen by more traditional fraud detection methodologies.

After a thorough assessment by our solution architects, the model was implemented using a cloud architecture. Special attention has been given to the secure processing of sensitive cloud data while ensuring low latency for real-time fraud detection.

We have designed, implemented and optimized an architecture consisting of multiple micro-services that have been replicated in three distinct environments: development, testing and production.

Big Data context on payment fraud prevention

The context of big data

The scope of data encompasses all data from the year 2020, during which our customer managed about 500 million transactions. Given this immense amount of data, the use of a distributed system for running pre-processing queries is needed.

Moreover, compared to other machine learning models, deep neural networks perform the best in this context, as they can learn from a huge amount of information.

Highly imbalanced classification

Highly imbalanced classification

The number of fraudulent transactions is much smaller than the number of genuine transactions. This implies that the classification will be between two highly imbalanced classes.

To solve this problem, Nextbit has adopted weighted sampling techniques and appropriate loss functions, both of which come out of recent academic research.

History of transactions

Learning from the past

To predict whether a transaction is fraudulent or genuine, the model uses information contained in more than 25 features. Some of them are categorical variables and are composed of hundreds of categories.

Despite this, the neural network manages to avoid the problem of high dimensionality by utilizing a proper initial layer. Moreover, to detect anomalous behaviours, which may indicate the presence of fraud, the neural network considers information on the history of transactions, which requires the use of an appropriate neural network for analysing sequential data.

towards the future

Automating fraud risk assessment

Financial institutions are adopting strategies to manage and prevent fraud with the long-term goal of making digital payments completely secure and reliable.

Nextbit’s artificial intelligence approach leverages vast amounts of available data and makes digital payments more secure by detecting patterns not seen by more traditional fraud detection methodologies.

Find out how Nextbit can empower your team today

Drop us a line at