Federated Transfer Learning for Cross-Institution Fraud Detection in Finance

Salge, Aryan Nagnath (2025) Federated Transfer Learning for Cross-Institution Fraud Detection in Finance. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (1MB) | Preview
Preview	PDF (Configuration Manual) Download (3MB) | Preview

Abstract

Financial fraud detection is a critical task for financial institutions, but privacy and regulatory constraints often prevent cross-institutional data sharing, limiting the effectiveness of machine learning models trained on isolated datasets. This study examines the application of Federated Transfer Learning (FTL) to facilitate collaborative fraud detection while maintaining data confidentiality. Three heterogeneous real-world fraud datasets, representing independent financial institutions with differing sizes, class distributions, and partially overlapping features, harmonized for joint training, were used to simulate a cross-institutional learning environment.

The proposed framework coordinates shared features, applies Principal Component Analysis (PCA) for dimensionality reduction, and trains a global neural network model using the Federated Averaging (FedAvg) algorithm [1] under differential privacy constraints. A fine-tuning step is then performed on each of the three clients to personalize the model to local data distributions. Performance is examined using Accuracy, Precision, Recall, and F1-Score, and compared against a centralized baseline model trained on aggregated data. Results show that FTL achieves accuracy levels above 98% across all clients, closely approaching centralized performance, while enabling privacy-preserving collaboration. Fine-tuning particularly benefits datapoor and imbalanced clients, improving accuracy from 57.4% to 99.4%, highlighting the potential use of knowledge transfer in a federated environment. This dramatic performance gain occurred despite Dataset 3 initially containing the most missing values and lowest fraud ratio, illustrating that FTL can successfully uplift low-quality, under-resourced clients through privacy-preserving collaboration.

Despite promising outcomes, limitations of the experiment include convergence instability on non-IID data, privacy guarantees, and a simulated deployment environment. Future research directions include integrating advanced aggregation methods such as FedProto [2] and adopting secure multiparty computation. This study provides an experimental proof-of-concept for FTL as a viable approach to collaborative fraud detection under strict privacy constraints.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Hamill, David UNSPECIFIED
Uncontrolled Keywords:	Federated Transfer Learning; Financial Fraud Detection; Cross-Institution Learning; Privacy Preservation; Centralized Baseline
Subjects:	H Social Sciences > HG Finance Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	03 Jul 2026 10:07
Last Modified:	03 Jul 2026 10:07
URI:	https://norma.ncirl.ie/id/eprint/9457

Actions (login required)

View Item