SynB2B-Fraud: A Synthetic Dataset for Fraud Detection in B2B Payment Networks

A fully synthetic B2B payment-fraud dataset — 100,000 buyer→supplier invoice transactions with five labeled fraud topologies. No real data is used, so it is freely redistributable under CC BY 4.0. DOI: 10.5281/zenodo.21668281 Companion to STREAM-BSG: A Streaming Graph Architecture for Real-Time Fraud Detection in B2B Payment Networks (IEEE InC4 2026). Files File Description synb2b_fraud.csv / synb2b_fraud.parquet The dataset (100,000 rows, 14 columns) synb2b_fraud_stats.json Summary statistics for this release generate_synb2b_fraud.py Seed-deterministic generator (regenerate at any size) SynB2B-Fraud_DATASHEET.md / .pdf Full datasheet (Datasheets for Datasets format) At a glance 100,000 transactions · 1,250 buyers · 3,198 suppliers · Jan 2024 – Jul 2025 1,500 fraud rows (1.5%), balanced 300 each across T1–T5 Labels: fraud_injected (0/1) and fraud_topology (T1–T5 / null) Fraud topologies: T1 vendor injection · T2 invoice cycling · T3 payment-term manipulation · T4 shell-supplier ring · T5 wire redirection (BEC)