The development process of a new product based on machine learning presented several unique challenges:

• Product development: Creating a product that anti-fraud analysts could easily appreciate, integrating advanced machine learning technologies.
• Academic foundations: Using the Politecnico di Milano research as a basis for developing an innovative product reimplementing complex algorithms for data analysis.
• On-premises environment: Facing the challenge of working with huge volumes of sensitive data, requiring an on-premises implementation due to the delicate nature of the information managed by financial institutions.

We developed and implemented a robust solution capable of efficiently analyzing vast volumes of data. Banksealer has evolved into a powerful semi-automated asset for financial analysts, enabling them to pinpoint high-risk transactions via a sophisticated ranking system grounded in "anomaly scores".

One of the hardest issues for companies developing machine learning products is bringing them to production.

Working on Banksealer, we faced some of the most challenging production-related problems: handling millions of transactions per day, deploying the system on-premises with high resource constraints, and implementing ad-hoc machine learning algorithms.

Banksealer is a semi-automatic system whose primary goal is to help financial analysts identify risky transactions. The algorithms rank transactions by "anomaly score" and smartly present the riskiest ones to the analyst.

The algorithms we used were developed in collaboration with Politecnico di Milano by Professor Michele Carminati. They are optimized to effectively identify fraud attacks and provide insights about the attack to the analyst. That's why we could not use out-of-the-box ones. The algorithms were developed in Scala programming language in collaboration with Politecnico di Milano and run against the datasets used in the research papers to check their detection performances.

Banksealer's data flow is very complex: it needs to integrate with many external systems, such as ticket managers, dashboards, search engines (ELK), and other fraud detection tools.

We used a micro-service architecture to obtain a scalable architecture and easily evolve single components. We adopted Apache Kafka (a message queue system) to flawlessly manage the data flow between our services and achieve near real-time performances. All the services are independent and stateless. The system can be evolved without the risk of compromising the availability of the existing data.

The system is installed on-premises in an environment without an internet connection. To quickly install and update our services, we dockerized every part of our software. The system can be installed easily both as a virtual appliance or on any operating system that supports docker.

The architecture is so flexible that different installation configurations are available. For instance, the database can be embedded in the appliance or managed externally if the customer has enough resources. We also leveraged docker to supervise the processes and provide fault tolerance to the system.

Although we implement complex machine learning algorithms, the system has light hardware requirements. It doesn't necessarily require GPUs, and it runs on standard machines. We quickly adapted Banksealer to work with different data, and we have successfully deployed it on both single-instance and PaaS configurations.

The success of Banksealer demonstrates the effectiveness of close collaboration between the academic and corporate worlds in developing innovative solutions to combat financial fraud. Our ability to constantly adapt and optimize our approach to emerging cybersecurity challenges underlines our commitment to providing effective and reliable tools for protecting financial data.