Source-linked AI summary

Realistic Synthetic Financial Transactions for Anti-Money Laundering Models

Erik Altman, Jovan Blanuša, Luc von Niederhäusern, Béni Egressy, Andreea Anghel, Kubilay Atasu

arXiv:2306.16424v3cs.AIcs.LGq-fin.CP

TL;DR

Money-laundering detection lacks broadly available real transaction data and complete labels, while prior synthetic AML efforts have important shortcomings. The paper introduces AMLworld, a calibrated multi-agent generator and public dataset suite with complete laundering tags, and uses it to benchmark machine-learning models. The datasets support model comparison and exploration of laundering patterns and scenarios that real data cannot fully expose.

  • Problem

    Real financial transaction data is generally unavailable, and laundering labels are incomplete because many laundering transactions go undetected; prior synthetic AML efforts also have major shortcomings.

  • Method

    AMLworld uses a multi-agent virtual world to generate realistic synthetic financial transactions and propagate laundering tags through transactions, including co-mingled funds.

  • Results

    AMLworld data matches real-world data in key regards, provides complete laundering tagging, and supports baseline comparisons of machine-learning models.

  • Takeaways & Limitations

    The public datasets enable precise model comparison and generation of what-if laundering patterns not yet observed in real data.

  • Takeaways & Limitations

    The synthetic data is static, making it harder for launderers to change behavior specifically to evade detection than if detection algorithms were published.

Abstract

from arXiv · show

With the widespread digitization of finance and the increasing popularity of cryptocurrencies, the sophistication of fraud schemes devised by cybercriminals is growing. Money laundering -- the movement of illicit funds to conceal their origins -- can cross bank and national boundaries, producing complex transaction patterns. The UN estimates 2-5\% of global GDP or \$0.8 - \$2.0 trillion dollars are laundered globally each year. Unfortunately, real data to train machine learning models to detect laundering is generally not available, and previous synthetic data generators have had significant shortcomings. A realistic, standardized, publicly-available benchmark is needed for comparing models and for the advancement of the area. To this end, this paper contributes a synthetic financial transaction dataset generator and a set of synthetically generated AML (Anti-Money Laundering) datasets. We have calibrated this agent-based generator to match real transactions as closely as possible and made the datasets public. We describe the generator in detail and demonstrate how the datasets generated can help compare different machine learning models in terms of their AML abilities. In a key way, using synthetic data in these comparisons can be even better than using real data: the ground truth labels are complete, whilst many laundering transactions in real data are never detected.

1 Introduction

Money laundering creates complex transaction patterns, while legal, privacy, and labeling barriers limit access to real training data. The paper addresses this gap with a realistic synthetic generator, public AML datasets, and baseline experiments.

  • Motivation: Real financial data is generally unavailable, and laundering labels are incomplete because many illicit transactions go undetected.These constraints make it difficult to build and compare laundering-detection models using real data.
  • Data Representation: Synthetic financial transactions can be represented in tables or graphs, with graphs making account connectivity and interaction patterns explicit.Tabular records include source and destination accounts, amounts, timestamps, currencies, and payment types.
  • Motivation: Financial transaction graphs expose account connectivity and complex interaction patterns, including cycles that strongly indicate laundering.Transactions can also involve cryptocurrency mixing and shuffling schemes represented as subgraph structures.
  • Contributions: AMLworld generates labeled synthetic transactions in a multi-agent virtual world containing criminal agents with illicit income to launder.The generator provides perfect information about laundering transactions and the specific patterns used in each instance.
  • Contributions: The paper releases realistic, standardized AML datasets spanning multiple sizes and difficulties for developing and benchmarking detection models.The release includes initial GNN and GBT experiments as baselines, with open-source GNN code and reproducible GBT results.

2 Related Work

Prior AML research is constrained by unavailable real data, incomplete labels, and synthetic datasets that may require real data or omit important attributes. This work positions AMLworld as a more detailed, scalable synthetic alternative for standardized comparison.

  • Data Availability: Real transaction data is rarely available, and even anonymized data lacks known ground truth because laundering often goes undetected.These constraints limit reliable evaluation of AML models.
  • Prior Studies: Prior real-data studies are narrow in access or pattern coverage, including one private dataset that searched for only two laundering motifs without guaranteeing complete detection.Another study used transaction summaries to assess customer laundering risk rather than providing broadly reusable transaction data.
  • Synthetic AML Data: Existing synthetic AML efforts may require real data, lack authoritative laundering labels, or omit important attributes of real transactions.These shortcomings motivate a generator that does not depend on instances of real data being imitated.
  • Agent-Based Simulations: Agent-based AML simulations differ in modeling detail and scale, which matters because illicit transactions form a tiny minority class.Table 1 compares AMLworld with prominent previous synthetic AML efforts.
  • Detection Models: Graph Neural Networks are designed for relational data and have been applied to money laundering, phishing, and fraudulent-transaction detection.This establishes GNNs as relevant model families for transaction-network AML tasks.

3 Synthetic Datasets

The paper introduces AMLworld, a fully synthetic multi-agent financial world that generates realistic, labeled AML datasets for benchmarking and what-if analysis.

  • 3.3 What is new in our AMLworld generator?: AMLSim models layering, whereas AMLworld covers placement, layering, and integration across nine criminal-income sources.AMLworld also tracks illicit funds from their origins through later financial activity.
  • 3.4 The Virtual World Model: AMLSworld builds a multi-agent virtual world of banks, individuals, and companies whose ordinary activities generate financial transfers alongside laundering behavior.Agents buy items, pay salaries and pensions, transfer funds, and may engage in activities such as smuggling or natural laundering through supplies and employee payments.
  • 3.5 Public Availability of Synthetic Data: The generator produces billions of labeled transactions across currencies including dollars, euros, yuan, yen, and bitcoin.Each transaction is labeled as laundering or not, and the generator supports varied dataset sizes and illicit-transaction ratios.
  • 3.4 The Virtual World Model: AMLworld uses statistical distributions rather than real transaction seeds and propagates laundering labels through transfers, including commingled funds.This transitive tracking provides complete laundering labels that are difficult to obtain from real-world data.
  • 3.5 Public Availability of Synthetic Data: The public HI and LI datasets span small, medium, and large scales, with large datasets containing roughly 175M–180M transactions.LI-Large contains 176M transactions over slightly more than three months, and its transaction and payment-format statistics roughly align with Federal Reserve data.
  • 3.5 Public Availability of Synthetic Data: 1 in 21,440 LI-Large transactions belongs to Gather-Scatter or Scatter-Gather patterns, closely matching a real-bank comparison of 1 in 21,000.LI-Large contains 8,212 combined transactions in these patterns, while the real-bank study reports 855 laundering motifs among 180M transactions.

4 Performance Evaluation of Machine Learning Models

The evaluation benchmarks gradient-boosted trees and graph neural networks on imbalanced synthetic AML datasets, emphasizing minority-class F1. Shared models and cross-dataset pretraining improve performance, while lower-illicit-ratio datasets remain more challenging.

  • The experiments use temporal 60-20-20 train-validation-test splits, edge-classification GNNs with neighborhood sampling, and account identifiers excluded as features.GNN validation and test graphs retain earlier transactions so models can identify patterns in unseen transaction batches.
  • Message-passing GNNs and GFP-enhanced gradient boosting effectively detect laundering transactions, while PNA performs near-parity with gradient-boosting baselines without handcrafted features.The evaluation emphasizes minority-class F1 because the datasets are naturally imbalanced.
  • LI datasets are harder than HI datasets because laundering is less frequent and its patterns span longer time periods.The models therefore weight minority-class predictions more heavily in the loss function.
  • Models pretrained on HI-Medium and HI-Large achieve higher LI-dataset F1 scores than XGBoost models trained directly on the corresponding LI datasets.Table 3 reports minority-class F1 scores for models trained on HI datasets and evaluated on LI datasets.
  • Shared models achieve higher F1 scores, with shared financial transaction graphs producing larger gains than shared models alone.The authors identify differentially private model and topology sharing as potential ways to obtain these improvements.

5 Future Work and Research Avenues

The paper identifies cross-bank analysis, complex graph-pattern detection, transfer learning, efficiency, and generator stability as future research opportunities enabled or motivated by AMLworld.

  • Synthetic data enables analysis across multiple banks, motivating privacy-preserving machine learning and graph topology-sharing techniques.
  • Future AML models should detect complex laundering patterns in graph datasets across single-bank and cross-bank settings.
  • Initial experiments in Table 3 demonstrate that transfer learning from synthetic cross-bank data is viable before fine-tuning with real-world data.
  • Efficient AML algorithms must account for time, memory, CPU, and GPU use because transaction-analysis requirements differ across domains.
  • Agent-based generation can be unstable because minor parameter changes may significantly alter aggregate outcomes such as the laundering rate.

6 Conclusions

The paper presents AMLworld as a detailed multi-agent generator producing realistic, perfectly tagged laundering data and public datasets for AML model development and benchmarking.

  • AMLworld combines a detailed multi-agent virtual world with synthetic financial data labeled for money laundering.
  • Generated data matches real-world data in key regards while providing perfect laundering tags that real financial data generally lacks.
  • Six AMLworld datasets are publicly available on Kaggle, supporting development and benchmarking of money laundering detection models.
  • Initial evaluation found GNNs and GBTs effective for identifying laundering transactions, with GNNs often competitive without feature engineering and GBTs requiring it for complex patterns.
  • AMLworld also supports generating models for what-if laundering patterns not yet observed in real data.

7 Ethics

The ethics discussion weighs AMLworld’s benefits for improving detection against the possibility of misuse, emphasizing that releasing static data limits adaptive evasion compared with releasing detection algorithms or generator code.

  • Most laundering activity goes undetected, motivating data release aimed at narrowing the detection gap with credit card fraud.
  • Unlike real laundering data, AMLworld provides ground-truth labels because real-world laundering lacks reliable ground truth and consequently produces missed fraud and false negatives.
  • AMLworld’s synthetic data is intended to improve laundering detection despite the possibility of limited benefit to launderers.
  • Publishing a static dataset makes adaptive evasion harder than publishing a bank’s detection algorithms, which launderers could use to change behavior.
  • The authors release data rather than generator code because source code would make it easier to tweak outputs and test them against detection algorithms.
  • Additional ethics observations are provided in Appendix B.

Supplementary Material

The supplementary material documents AMLworld’s public datasets, their HI/LI groupings and scales, and statistics showing alignment with real transaction, tax, and demographic distributions.

  • AMLworld provides six public datasets organized into higher- and lower-illicit groups, each further divided into small, medium, and large sizes.Large datasets contain 175M–180M transactions.
  • Supplementary tables describe laundering-pattern node counts, pattern occurrences, laundering rates, transaction formats, public dataset statistics, and hyperparameter configurations.
  • Transaction-rate histograms and transaction-format distributions roughly align with U.S. Federal Reserve data.
  • Laundering transactions comprise eight standard laundering patterns plus integration transactions disguised as activities such as payroll or company supplies.
  • Salary and pension amounts are based on U.S. Internal Revenue Service data to support accurate modeling of transaction sizes and frequency.

B Ethical Use of the Data

The paper frames the synthetic data as useful for benchmarking and model development while documenting how graph features and baselines are constructed without data leakage.

  • B Ethical Use of the Data: Synthetic data are intended for benchmarking and improving money-laundering detection models, with performance interpreted cautiously because transfer to real data may not be one-to-one.The authors also note possible use for pretraining before fine-tuning on real data.
  • B Ethical Use of the Data: LightGBM and XGBoost hyperparameters are tuned with successive halving, which progressively evaluates fewer configurations on larger training fractions.The method begins with randomly sampled configurations and retains the best x0/η configurations for later rounds.
  • B Ethical Use of the Data: GNN hyperparameters are selected through two rounds of random sampling over layers, embedding size, learning rate, dropout, and minority-class weighting.Final models use the best validation configuration with different random seeds.
  • B Ethical Use of the Data: The Graph Feature Preprocessor streams temporal edges through an in-memory dynamic graph to add features such as scatter-gather patterns, cycles, and vertex statistics.The experiments use batch size 128, specified time windows, and timestamp ordering so features depend only on past data.

E Additional GNN Experiments

Additional GNN experiments report computational costs and finer-grained precision and recall results for the baseline models.

  • E Additional GNN Experiments: The GNN implementations are publicly available through supplementary material and GitHub under an Apache License.They use PyTorch Geometric 2.3.1 and PyTorch 2.0.1.
  • E Additional GNN Experiments: GNN baseline training and hyperparameter optimization required approximately 1000 GPU hours on Nvidia Tesla V100 GPUs.Runtime measurements cover the AML Small and Medium datasets, with models fixed at two layers and hidden size 64.
  • E Additional GNN Experiments: The additional evaluation separates minority-class recall, precision, and F1, clarifying the trade-off between finding laundering instances and avoiding incorrect predictions.Recall, precision, and F1 results are reported in separate tables, with precision-recall curves provided as examples.

F Additional GBT Experiments

Additional GBT experiments examine precision-recall behavior and the effect of sharing transaction graphs and models across banks.

  • F Additional GBT Experiments: XGBoost precision-recall curves using GFP features show that increasing recall beyond the 0.5-threshold point sharply degrades precision.The curves correspond to the GFP+XGBoost results in Table 2.
  • F Additional GBT Experiments: Precision-recall curves are also provided for the best-performing PNA models across all AML datasets, with red dots marking F1 at threshold 0.5.The supplied figure identifies the curves and threshold but does not state an outcome comparison.
  • F Additional GBT Experiments: 20.8% and 22.1% average minority-class F1 for LI-Medium and LI-Large, respectively, when banks share the transaction graph and global model.The corresponding private-graph, private-model averages are 4.9% and 8.7%, showing the shared setting performs better on these lower-illicit-ratio datasets.
  • F Additional GBT Experiments: The per-bank LI experiment is presented for LI-Medium and LI-Large datasets, which contain fewer illicit transactions than the corresponding HI datasets.This lower illicit ratio makes local-bank modeling more challenging.

G Datasheet

The datasheet describes six synthetic transaction-network datasets created for financial-crime detection, with simulated labels and recommended use conditions.

  • G Datasheet: The datasets were created to test, develop, and improve machine-learning models for detecting money laundering transactions.They are distributed through Kaggle, which serves as the single source of distribution and maintenance.
  • G Datasheet: Each dataset is one synthetic financial transaction graph containing 5M to 180M transactions, with nodes as accounts or entities and directed edges as transactions.Edge features include amount, currency, and transaction type, and the information is entirely simulated.
  • G Datasheet: The datasets contain no personally identifiable, confidential, or offensive content because all account and transaction details are simulated.The datasheet states that the dataset is self-contained.
  • G Datasheet: Transactions have labels, no information is missing, and recommended data splits are provided.The datasets can generate arbitrary numbers of transactions rather than representing a fixed sample from a larger population.
  • G Datasheet: The authors report no known errors, noise, or redundancies in the datasets.This statement is qualified as being based on the authors’ knowledge.
Loading 2306.16424v3…