The Future of Fraud Detection in Financial Institutions: How Federated Learning (FL) Solves Key Challenges

Federated Learning

Financial Crime

April 2, 2025

In Part 1, we discussed the challenges associated with centralized machine learning (ML) systems in the financial sector, particularly their limitations in addressing the diverse nature of financial crimes like money laundering. One of the key issues is that each bank typically only encounters a narrow set of fraudulent activities, such as large cash deposits or transactions with entities on sanctions lists. This siloed data means that each institution lacks a broader view of money laundering tactics, which hampers the effectiveness of fraud detection models.

In this post, we will explore how Federated Learning (FL) can overcome these challenges by enabling financial institutions to collaboratively build fraud detection models without compromising sensitive data privacy. Through FL, banks can work together to improve their fraud detection capabilities while ensuring compliance with data privacy regulations and enhancing their ability to detect a wide variety of fraudulent activities.

Why Traditional Machine Learning Models Struggle

Traditional machine learning models face significant limitations in detecting a broad range of fraudulent transactions due to two key factors:

Limited Fraud Coverage: Banks can only detect the types of fraud they have already encountered and for which they have existing models. Since fraud detection models are trained on historical data, each bank’s model is narrowly focused on specific fraud patterns, such as large cash deposits or transactions involving sanctioned entities. This means they can only catch scenarios they already have models for, leaving other types of fraud undetected.

Insufficient Positive Labels for Training: Fraud detection models rely on labeled data, but fraudulent transactions are relatively rare compared to legitimate ones. As a result, banks often have very few positive fraud labels available to train their models. This imbalance creates models that are less effective in identifying emerging or less common fraud patterns, limiting their ability to respond to new threats.

This is where Federated Learning (FL) offers a significant advantage. FL enables banks to collaborate on fraud detection without sharing raw data, allowing them to leverage a much broader range of fraud scenarios and more diverse training data. By pooling insights across institutions, FL can help create more robust models that generalize better across a wider variety of fraud types and make better use of the limited positive labels available.

How Federated Learning Works: A Decentralized Approach

Federated Learning provides a solution to these challenges by enabling multiple institutions to collaborate on training a machine learning model without ever sharing their raw data. Here’s how it works:

Decentralized Model Training: Each institution trains a model locally on its own data, ensuring that sensitive customer information remains private and secure.

Model Updates: Instead of sharing raw data, each institution sends model updates (i.e., the weights of the neural network) to a central server.

Global Model Aggregation: The central server aggregates these updates and creates a global model, which is then sent back to all institutions for further refinement.

Privacy Preservation: Since no raw data is exchanged between institutions, privacy is maintained, ensuring compliance with data privacy regulations.

Iterative Improvement: This process repeats multiple times, with the model improving over each round of training, leading to a more accurate and robust fraud detection system.

Federated Learning allows financial institutions to collaborate and create a more comprehensive fraud detection system without the need to share raw transaction data, ensuring privacy and regulatory compliance.

Demonstrating the Power of Federated Learning: The Soteria Initiative Demo

To illustrate the potential of FL in the financial sector, we developed a demo in collaboration with the Soteria Initiative, an organization focused on fighting financial crime through data standards and innovation in the financial sector. In this demo, we simulated four banks working together to detect various types of fraudulent transactions, including:

Large cash deposits

Transactions to Russia

Transactions involving watchlisted entities

Transactions linked to entities with many Ultimate Beneficial Owners (UBOs)

We used a common transaction format provided by the Soteria Initiative to ensure interoperability. You can explore the transaction format here: Soteria CoreData Transaction Format.

Each simulated institution trained a model on one of these fraud scenarios, and then we evaluated the ability of each model to detect a single fraud scenario versus multiple scenarios. The results were striking:

Traditional Model Performance: When trained on a single fraud scenario, the AUC (Area Under the Curve) on the test set was 0.730. However, when tested on multiple fraud scenarios (4 types of fraud), the AUC dropped significantly to 0.545, indicating a performance close to random.

Federated Learning Performance: The FL model demonstrated significant improvement. After just one round of training, it achieved an AUC of 0.9083 when tested on all four fraud scenarios, showing its ability to generalize across multiple types of fraud.

The ability of the FL model to improve fraud detection performance across a broader range of scenarios highlights its superior capabilities compared to traditional models that rely on centralized data.

You can access the full demo and code on GitHub here: Soteria FL Demo.

Key Benefits of Federated Learning for Financial Institutions

Federated Learning offers several significant advantages over traditional centralized machine learning models for fraud detection:

Improved Fraud Detection: By learning from a wide range of fraud scenarios across multiple institutions, FL helps improve the accuracy and robustness of fraud detection models. As demonstrated in the Soteria Initiative demo, FL can detect fraud more effectively across diverse scenarios, leading to higher detection rates and reduced false positives.

Privacy Preservation: FL allows financial institutions to collaborate on fraud detection without sharing sensitive transaction data. This decentralized approach respects privacy laws and avoids the risks associated with data breaches or regulatory non-compliance.

Cost Efficiency: By pooling intelligence from multiple institutions, FL reduces the need for each institution to build and maintain its own fraud detection systems, lowering operational costs. Banks can leverage the collective intelligence of the system while maintaining data sovereignty.

Regulatory Compliance: The decentralized nature of Federated Learning ensures that financial institutions comply with stringent data protection regulations like GDPR, as no raw data is exchanged between institutions.

What’s in it for Banks and Regulators?

For financial institutions, Federated Learning represents an opportunity to improve fraud detection capabilities at a reduced cost while ensuring privacy and compliance. Rather than working in isolation, banks can collaborate with others to create more comprehensive fraud detection models, addressing a wider array of fraud patterns without violating customer privacy.

From a regulatory perspective, FL offers a way to improve the financial industry’s collective ability to combat fraud without sacrificing privacy. As regulators continue to push for stronger anti-money laundering (AML) measures, supporting the adoption of Federated Learning could help improve overall fraud detection while adhering to data protection laws.

Conclusion: The Power of Federated Learning in the Fight Against Financial Crime

Federated Learning represents a paradigm shift in how financial institutions can collaborate to detect and prevent fraud. By enabling the sharing of model insights without compromising sensitive customer data, FL provides a powerful, decentralized approach to building more robust fraud detection systems. The results from the Soteria Initiative demo demonstrate the real-world effectiveness of FL in detecting multiple fraud scenarios, outperforming traditional models that rely on centralized data.

As the financial industry continues to evolve, embracing innovative technologies like Federated Learning will be crucial for staying ahead of fraudsters while respecting privacy and regulatory requirements. Financial institutions and regulators alike should consider the potential of FL in enhancing the fight against financial crime.