Adversarial Training for Fraud Detection in Mobile Finance in Bangladesh: A Hybrid Approach Using Machine Learning
Year: 2025
Type: conference-paper
Abstract: As mobile money services like bKash and Nagad gain traction in Bangladesh, fraud detection has become a critical issue. Traditional machine learning methods, such as Logistic Regression, Random Forest, and XGBoost, are ineffective when transaction-level data is exploited for nefarious ends. This research introduces a hybrid fraud detection system that combines adversarial machine learning techniques with conventional methods. We empirically examine the efficacy of shallow models in comparison to the Multi-Layer Perceptron (MLP) model utilizing adversarial training techniques such as FGSM and PGD. Our findings indicate that whereas shallow models can perform well on pristine data, they lack resilience against attacks. Conversely, adversarially trained MLP models demonstrate resilience, affirming that adversarial training is essential for developing fraud detection systems capable of withstanding real-world attacks in the mobile financial services sector of Bangladesh.
Institution: World University of Bangladesh
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7224936202