FraudGraph: Transaction & Graph-based Fraud Detection Toolkit

FraudGraph is an open-source toolkit to build transaction networks from event logs and run scalable graph algorithms — community detection, link analysis, and suspicious subgraph scoring — to surface collusive fraud rings.

Why this matters

Large fraud schemes are rarely isolated. They behave like networks: coordinated accounts, shared devices, and recycled payment instruments. Finding the ring (not just the bad individual) helps platforms, investigators, and researchers stop coordinated attacks and reduce harm at scale.

Features

• ETL → Graph: From flat events to a heterogeneous network (users, devices, cards, IPs) with typed edges.
• Algorithms:
  • Louvain community detection
  • HITS / PageRank-style authority-hub analysis
  • Suspicious subgraph scoring (density, triads, shared-identifier overlap, burstiness)
• Scalable paths:
  • Prototype with NetworkX
  • Optional connector to Neo4j for larger graphs (Cypher + APOC friendly)
• Reproducible demos: Jupyter notebook and an end-to-end example with a simulated collusive ring.
• Ethics & Safety by design: Clear docs on anonymization, differential risk of false positives, and reproducibility.

Quickstart

# 1) Install (recommended: Python 3.10+)
python3 -m venv .venv && source .venv/bin/activate
pip3 install -r requirements.txt  # or: pip3 install -e .

# 2) Run the simulated example
python3 examples/simulate_ring.py --out data/sim_transactions.csv
python3 examples/run_pipeline.py --transactions data/sim_transactions.csv --out reports

# 3) Explore results
# Open the generated report:
open reports/report.md
# (Optional) launch the notebook for visualization:
# jupyter lab notebooks/demo_ring.ipynb

Data

• Synthetic by default (provided script) to avoid handling real PII.
• You can also point the ETL to public research datasets that include transaction-like edges (e.g., anonymized credit-card fraud samples, synthetic fintech traces). Always ensure proper licensing and anonymization.

Project layout

fraudgraph-toolkit/
├─ src/fraudgraph/
│  ├─ etl.py                  # Build graphs from CSV/JSONL transactions
│  ├─ algorithms.py           # Louvain, HITS, suspicious subgraph scoring
│  └─ connectors/
│     └─ neo4j_connector.py   # Example writer/reader to Neo4j
├─ examples/
│  ├─ simulate_ring.py        # Generate synthetic ring & benign background
│  └─ run_pipeline.py         # CLI: ETL → algorithms → markdown report
├─ notebooks/
│  └─ demo_ring.ipynb         # Walkthrough & visualizations
├─ docs/
│  └─ WHITEPAPER.md           # Short, non-marketing technical summary
├─ data/                      # (gitignored) place local inputs here
└─ README.md

Reproducing the case study

1. Generate or supply a CSV of transactions with columns:
   • timestamp, sender, receiver, amount, device_id, ip, instrument_id
2. Run the pipeline (see above) to create a graph and produce a lightweight report with top communities and scores.
3. Inspect the notebook for visual exploration and parameter tuning.

Societal impact

This toolkit supports research and engineering work to reduce large-scale coordinated financial scams by highlighting suspicious clusters for deeper human review. It does not perform automated bans; it provides signals to assist analysts and investigators.

Contributing

Issues and PRs are welcome! See docs/WHITEPAPER.md for the threat model and CONTRIBUTING notes inside the README.

License

MIT (see LICENSE).