AMLNet

AMLNet is a synthetic anti-money laundering benchmark dataset created for machine learning evaluation. This Version 2.0 release is associated with the paper: "AMLNet: A Knowledge-Guided Synthetic Benchmark for Machine Learning Evaluation in Anti-Money Laundering" The dataset contains a fixed synthetic benchmark instance generated using the AMLNet framework. It includes approximately 1.09 million transactions over a 195-day simulation period, from 13 October 2025 to 27 April 2026. The benchmark contains 1,411 suspicious transactions, corresponding to a suspicious rate of approximately 0.13%. The dataset is fully synthetic. The accounts, transactions, timestamps, locations, balances, metadata, labels, and customer activity patterns do not correspond to real customers, real institutions, or real banking activity. AMLNet was designed to support machine learning experiments under rare-event anti-money laundering conditions. The dataset includes ordinary transactions and suspicious transaction sequences representing structuring, layering, and integration patterns. Suspicious activity is embedded within ordinary account activity to make the detection task more realistic and non-trivial. CONTENTS This release includes: - the fixed AMLNet Version 2.0 synthetic transaction dataset;- transaction labels;- laundering typology labels;- transaction metadata;- dataset documentation and column descriptions. Evaluation scripts are not included in the current release, but can be made available to reviewers on request for verification. They will be added to the Zenodo record in a subsequent version after publication. The associated manuscript provides the generator pseudocode, main configuration details, demographic assumptions, regulatory constraints, laundering typologies, timing rules, routing rules, split protocol, and evaluation protocol. The AMLNet generator source code is not included in this release due to security concerns. Requests for access to the generator source code can be considered for legitimate research purposes, subject to identity verification, institutional affiliation, and appropriate use conditions. DATA FORMAT The main dataset is provided as a CSV file with the following columns: - step: Sequential simulation step or transaction index.- type: Transaction type, such as BPAY, CASH_OUT, DEBIT, EFTPOS, NPP, OSKO, PAYMENT, or TRANSFER.- amount: Transaction amount in Australian dollars.- category: Transaction category, such as housing, food, transport, recreation, healthcare, education, utilities, shell company, property investment, cryptocurrency, or other.- nameOrig: Originating account or customer identifier.- nameDest: Destination account, customer, or merchant identifier.- oldbalanceOrg: Originating account balance before the transaction.- newbalanceOrig: Originating account balance after the transaction.- isFraud: Binary label used for suspicious/fraudulent transaction detection.- isMoneyLaundering: Binary AML label, where 1 indicates suspicious money laundering activity and 0 indicates ordinary activity.- laundering_typology: Laundering typology label. Values include normal, structuring, layering, and integration.- metadata: JSON-style metadata containing timestamp, location, device information, payment method, risk indicators, and typology-specific details where applicable.- fraud_probability: Model-generated or risk-score field used in selected experiments. This field may be empty for some records.- hour: Hour of transaction.- day_of_week: Day of week.- day_of_month: Day of month.- month: Month number. LABELS The dataset includes two binary label fields: - isFraud: Binary suspicious/fraudulent transaction label.- isMoneyLaundering: Binary anti-money laundering label. The laundering_typology column provides the suspicious activity type for labeled suspicious transactions. Values include: - normal- structuring- layering- integration DATASET STATISTICS - Total transactions: approximately 1.09 million- Suspicious transactions: 1,411- Suspicious transaction rate: approximately 0.13%- Simulation period: 195 days- Simulation dates: 13 October 2025 to 27 April 2026- Generated customer accounts: 10,000- Observed graph nodes: 11,000, including customer and merchant nodes- Transaction types: BPAY, CASH_OUT, DEBIT, EFTPOS, NPP, OSKO, PAYMENT, and TRANSFER- Payment identifiers: BSB_Account, CardNumber, and PayID- Laundering typologies: structuring, layering, and integration- Geographic setting: Australian synthetic banking context SPLIT AND EVALUATION PROTOCOL The associated manuscript describes the split protocol and evaluation protocol used in the reported experiments. For transaction-level experiments, transactions are ordered chronologically to reduce temporal leakage. For node-level graph experiments, the dataset is converted into an account/entity graph, and node features are computed from aggregated transaction statistics. The manuscript describes the chronological train, validation, and test protocol used for model evaluation. Evaluation scripts will be added to this Zenodo record after publication of the associated paper. SYNTHETIC DATA NOTICE This dataset is fully synthetic. It does not contain real customer records, real account information, real banking transactions, real IP addresses, or real financial institution data. All customer identifiers, merchant identifiers, balances, timestamps, locations, transaction patterns, labels, and metadata values were generated for research purposes. The dataset should not be interpreted as a sample of actual banking activity. USAGE This dataset can be used for: - anti-money laundering machine learning research;- rare-event classification experiments;- transaction-level suspicious activity detection;- account-level graph classification;- feature ablation studies;- explainability experiments;- synthetic-to-real transfer learning research;- educational and academic research on financial crime detection. LICENSE This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International License. You may share and adapt the dataset for non-commercial purposes, provided that appropriate credit is given. Commercial use is not permitted without prior permission.