•  
  •  
 

Document Type

Original Article

Abstract

Financial fraud detection is difficult because fraudulent transactions are rare, unevenly distributed, and sensitive. This study evaluates a lightweight multilayer perceptron under Federated Averaging (FedAvg), Federated Proximal optimization (FedProx), and Federated Adam (FedAdam), using three seeds and validation-selected decision thresholds. Corrected R4 experiments cover the Credit Card Fraud Detection benchmark, the larger Credit Card Transactions dataset, PaySim as a stand-alone synthetic stress test, and the Bank Account Fraud (BAF) Dataset Suite. On Credit Card, FedProx with 20 clients and alpha = 0.5 achieved F1 = 0.789 +/- 0.010. On Transactions, the best tested federated setting used 10 clients and achieved F1 = 0.347 +/- 0.027, retaining 64.3% of the centralized MLP F1 and 43.1% of the centralized HistGradientBoosting F1. The main-file PaySim run achieved F1 = 0.462 +/- 0.060, but its independent internal split precludes direct numerical comparison with the other datasets. On BAF Base, FedProx achieved F1 = 0.215 +/- 0.015, whereas Variant V collapsed to F1 = 0.039 +/- 0.012. Matched FedProx-FedAvg comparisons did not support general FedProx superiority. Eight-bit upload quantization preserved utility closely while reducing the calculated total payload by approximately 37.5%. These results emphasize threshold behavior, client fragmentation, dataset sensitivity, and payload-level communication rather than a single-model superiority claim.

Receive Date

24 May 2026

Revise Date

25 Aug 2026

Accept Date

06 Sep 2026

Publication Date

2026

Download

Share

COinS