Skip to content Skip to footer
NStarX Inc.
Get in Touch
NStarX Inc.
Let's Talk
Close

Optimizing Telecom Fraud Detection with Federated Learning

December 27, 2024

Problem Statement

The Telecom enterprise faced mounting challenges with fraud detection systems due to strict data sovereignty and privacy regulations across multiple geographies. Managing multiple Databricks instances was driving up costs, and the legacy logistic regression model struggled to handle complex fraud patterns. Furthermore, the existing system couldn’t scale effectively to adapt to emerging fraud techniques. The business needed a solution that could ensure data privacy while improving fraud detection accuracy and operational efficiency, leading to the adoption of Federated Learning.

Solution

NStarX revolutionized the Telecom enterprise’s fraud detection system by adopting Federated Learning (FL), an innovative decentralized approach to machine learning. The new system leveraged a multi-output deep learning model capable of predicting fraud scores, user classifications, and reason codes. Federated Learning allowed the model to train on decentralized data across multiple regions without compromising privacy or sharing sensitive data. Instead of moving data between systems, FL frameworks aggregated model updates locally, making the process both secure and efficient.

Advantages of Federated Learning:

Privacy Preservation

Federated Learning ensures that sensitive transaction data never leaves its local environment. The model is trained on the edge, keeping all data private and compliant with data sovereignty laws.

Reduced Data Movement

Since only model updates (not raw data) are shared between the clients and the central server, data movement is minimized, reducing network load and operational costs.

Scalable and Cost-Effective

With FL, the need for spinning up multiple centralized Databricks instances is eliminated. Model training happens at the local level, which reduces cloud infrastructure costs and scales effectively across geographies.

Technology

NStarX implemented Federated Learning through a combination of cutting-edge frameworks and tools:

Federated Learning Framework

We used the TensorFlow Federated (TFF) framework to implement federated learning. TFF allowed us to train machine learning models on decentralized data across different regions while keeping the data local. The federated averaging algorithm used in TFF ensured that model updates were aggregated securely and efficiently without violating privacy regulations.

Feast

A centralized Feature Store for managing features in a scalable, consistent way. Feast ensured that the same features were used across both training and production environments, and its real-time serving capabilities were critical for fraud detection.

MLFlow

Integrated with Feast, MLFlow was used for model management, versioning, and performance tracking, enabling continuous improvement of the fraud detection system and making it easier to monitor model’s performance over time and across different regions.

Business Outcomes

Fraud Detection Accuracy

Federated Learning improved the fraud detection accuracy by 35-40%, as the decentralized training allowed the model to capture more diverse patterns from data across different regions without compromising privacy.

Operational Cost Reduction

By leveraging Federated Learning and eliminating the need for duplicating data across data centers, the Telecom enterprise significantly reduced cloud infrastructure costs, saving over 30% in operational expenses.

Real-time Fraud Detection

With Feast providing real-time data serving and the integrated deep learning model, the system enabled faster fraud response times, minimizing losses from fraudulent transactions.

Data Sovereignty Compliance

The enterprise was able to maintain full compliance with regional data residency requirements while still benefitting from a collaborative, globally trained model through FL, ensuring legal adherence and maintaining trust with users.

Conclusion

By implementing Federated Learning (FL) with TensorFlow Federated, Feast, and MLFlow, NStarX empowered the Telecom enterprise to overcome critical data privacy and infrastructure challenges. The solution not only improved fraud detection accuracy but also optimized operational costs and ensured full compliance with data residency laws. The combination of decentralized model training, real-time data management, and secure collaboration across regions made FL an ideal solution for this Telecom enterprise’s evolving fraud detection needs.