CiferAI/Cifer-Fraud-Detection-Dataset-AF

📊 Cifer Fraud Detection Dataset

🧠 Overview

The Cifer-Fraud-Detection-Dataset-AF is a high-fidelity, fully synthetic dataset created to support the development and benchmarking of privacy-preserving, federated, and decentralized machine learning systems in financial fraud detection.

This dataset draws structural inspiration from the PaySim simulator, which was built using aggregated mobile money transaction data from a real financial provider operating in 14+ countries. Cifer extends this format by scaling it to 6 million samples, optimizing for federated learning environments, and validating performance against real-world datasets.

Accuracy Benchmark:

Cifer-trained models on this dataset reach 99.93% accuracy, benchmarked against real-world fraud datasets with 99.98% baseline accuracy—providing high-fidelity behavior for secure, distributed ML research.

---

⚙️ Generation Method

This dataset is entirely synthetic and was generated using Cifer’s internal simulation engine, trained to mimic patterns of financial behavior, agent dynamics, and fraud strategies typically observed in mobile money ecosystems.

• Based on the structure and simulation dynamics of PaySim
• Enhanced for multi-agent testing, federated partitioning, and async model training
• Includes realistic fraud flagging mechanisms and unbalanced label distributions

---

🧩 Data Structure

Column Name	Description
step	Unit of time (1 step = 1 hour); simulation spans 30 days (744 steps total)
type	Transaction type: CASH-IN, CASH-OUT, DEBIT, PAYMENT, TRANSFER
amount	Transaction value in simulated currency
nameOrig	Anonymized ID of sender
oldbalanceOrg	Sender’s balance before transaction
newbalanceOrig	Sender’s balance after transaction
nameDest	Anonymized ID of recipient
oldbalanceDest	Recipient’s balance before transaction (if applicable)
newbalanceDest	Recipient’s balance after transaction (if applicable)
isFraud	Binary flag: 1 if transaction is fraudulent
isFlaggedFraud	1 if transaction exceeds a flagged threshold (e.g. >200,000)

---

📁 File Organization

Total Rows: 6,000,000 Split into 4 folders/files for large-scale and federated learning scenarios:

• Cifer-Fraud-Detection-Dataset-AF-part-1-4.csv → 1.5M rows
• Cifer-Fraud-Detection-Dataset-AF-part-2-4.csv → 1.5M rows
• Cifer-Fraud-Detection-Dataset-AF-part-3-4.csv → 1.5M rows
• Cifer-Fraud-Detection-Dataset-AF-part-4-4.csv → 1.5M rows

Format: .csv (optionally .parquet or .json upon request)

---

✅ Key Features

• Fully synthetic and safe for public release
• Compatible with federated learning (cross-silo, async, or multi-agent)
• Ideal for privacy-preserving machine learning and robustness testing
• Benchmarkable against real-world fraud datasets
• Supports fairness evaluation via distribution-aware modeling

---

🔬 Use Cases

• Fraud detection benchmarking in decentralized AI systems
• Federated learning simulation (training, evaluation, aggregation)
• Model bias mitigation and fairness testing
• Multi-agent coordination and adversarial fraud modeling

---

📜 License

Apache 2.0 — freely usable with attribution

---

🧾 Attribution & Citation

This dataset was generated and extended by Cifer AI, building on structural principles introduced by:

E. A. Lopez-Rojas, A. Elmir, and S. Axelsson
PaySim: A financial mobile money simulator for fraud detection.
28th European Modeling and Simulation Symposium – EMSS 2016

---