PixGuard-Sim: A Deadline-Aware Testbed for Pix Fraud Detectors

• Cristhian Kapelinski UNIPAMPA https://orcid.org/0009-0005-5750-022X
• Diego Kreutz UNIPAMPA https://orcid.org/0000-0003-0830-0238

DOI: https://doi.org/10.5753/sbseg_estendido.2026.29851

Resumo

Brazilian regulation gives a paying institution about a second and a half to decide whether a suspect Pix transfer should be diverted for review before it settles, on a system used by 170 million people where settlement is irreversible. Yet every openly released payment-fraud dataset and baseline we surveyed is scored in batch, with no reference to that window. PixGuard-Sim is an open testbed that scores a detector against the window from the latency it measures, and releases a Pix-native generator with a settlement clock and two scenarios open data omits: multi-hop MED-2.0 refund tracing and payment under coercion. Scoring accuracy and timing together pulls them apart. The most accurate detectors we evaluate are cloud reasoning models that never saw a labeled event from this stream: the stronger finds 112 of 150 frauds where a random forest fitted to the data finds 93, and neither answers in time. Ranking by accuracy alone selects what cannot be deployed.

Referências

Altman, E., Blanuša, J., et al. (2023). Realistic synthetic financial transactions for anti-money laundering models. In NeurIPS Datasets & Benchmarks.

Aminian, G., Elliott, A., et al. (2025). FraudTransformer: Time-aware GPT for transaction fraud detection. In Workshop on AI for Financial Fraud Detection and Prevention at ACM ICAIF.

Braithwaite, J. (2024). ‘Authorized Push Payment’ bank fraud: What does an effective regulatory response look like? Journal of Financial Regulation, 10(2):174–193.

Chen, T. and Guestrin, C. (2016). XGBoost: A scalable tree boosting system. In KDD, pages 785–794.

Cheng, D., Zou, Y., et al. (2025). Graph neural networks for financial fraud detection: A review. Front. Comput. Sci., 19(9):199609.

Dal Pozzolo, A., Boracchi, G., et al. (2018). Credit card fraud detection: A realistic modeling and a novel learning strategy. IEEE TNNLS, 29(8):3784–3797.

Dal Pozzolo, A., Caelen, O., et al. (2014). Learned lessons in credit card fraud detection from a practitioner perspective. Expert Syst. Appl., 41(10):4915–4928.

Dou, Y., Liu, Z., et al. (2020). Enhancing graph neural network-based fraud detectors against camouflaged fraudsters. In CIKM.

Hamilton, W. L., Ying, R., and Leskovec, J. (2017). Inductive representation learning on large graphs. In NeurIPS.

Jiang, J., Li, Y., et al. (2022). Spade: A real-time fraud detection framework on evolving graphs. PVLDB, 16(3):461–469.

Johannessen, F. and Jullum, M. (2023). Finding money launderers using heterogeneous graph neural networks. arXiv:2307.13499.

Lopez-Rojas, E. A. and Axelsson, S. (2014). BankSim: A bank payments simulator for fraud detection research. In EMSS.

Lopez-Rojas, E. A., Elmir, A., and Axelsson, S. (2016). PaySim: A financial mobile money simulator for fraud detection. In EMSS, pages 249–255.

Lucas, Y., Portier, P.-E., et al. (2019). Dataset shift quantification for credit card fraud detection. IEEE AIKE, arXiv:1906.06977.

Pareja, A., Domeniconi, G., et al. (2020). EvolveGCN: Evolving graph convolutional networks for dynamic graphs. In AAAI, pages 5363–5370.

Pedregosa, F., Varoquaux, G., et al. (2011). Scikit-learn: Machine learning in Python. JMLR, 12:2825–2830.

Pizzolato, G. L., Lopes, B. M., et al. (2025). A taxonomy of pix fraud in brazil: Attack methodologies, ai-driven amplification, and defensive strategies.

van den Beukel, M., Rožanec, J. M., and Varbanescu, A.-L. (2026). Tide: A customisable dataset generator for anti-money laundering research. arXiv:2603.01863.

Weber, M., Chen, J., et al. (2018). Scalable graph learning for anti-money laundering: A first look. arXiv:1812.00076.

Weber, M., Domeniconi, G., et al. (2019). Anti-money laundering in Bitcoin: Experimenting with graph convolutional networks for financial forensics. KDD Workshop on Anomaly Detection in Finance, arXiv:1908.02591.

Yelleti, V. (2025). ROSFD: Robust online streaming fraud detection with resilience to concept drift in data streams. arXiv:2504.10229.