Fed-SMOTE: Differentially Private Federated Learning for Imbalanced Supply Chain Fraud Detection

Outlet Title

2026 IEEE 4th International Conference on Artificial Intelligence, Blockchain, and Internet of Things (AIBThings)

Document Type

Conference Proceeding

Publication Date

Fall 11-2026

Abstract

Due to the low frequency of fraudulent transactions in the supply chain, the inherently imbalanced nature of these datasets, and the privacy-sensitive data of the supply chain organizations, it is challenging to implement fraud detection systems using the centralized method or standard federated learning approaches. In this paper, we present a novel federated learning framework that leverages the advantages of both federated and SMOTE methods, called Federated Synthetic Minority Oversampling Technique (Fed-SMOTE). Each client calculates the statistics of the minority class of the supply chain transactions, applies the Laplacian mechanism to ensure the differential privacy of these statistics, and uses the statistics of the majority class calculated globally to generate synthetic transactions using the Borderline-SMOTE algorithm. These synthetic transactions are combined with the original transactions of each supply chain client to train an XGBoost classifier. The XGBoost classifiers from each client are combined using a performance weighted aggregation algorithm to create a final fraud detection model. Experiments on the DataCo Smart Supply Chain dataset with 180,519 transactions and 2.25% fraudulent transactions show that the Fed-SMOTE framework achieves an F1-score of 0.686 ± 0.004 and 99.5% recall with 10 random seeds. These results are comparable with the centralized SMOTE-based methods but maintain the benefits of data locality, differential privacy with ε=1.0, and perform well with non-IID datasets over a range of 2 to 20 clients with 86.7% communication cost reduction compared to the centralized method of sharing supply chain data. Thus, the results of this experiment demonstrate that SMOTE-based federated learning with differential privacy can provide a privacy-preserving, communication-efficient, and decision-oriented solution for fraud detection in supply chain environments.

Share

COinS