Open-Source Reference Implementations for AI-Enabled Payment Security, Fraud Detection, and AML/CFT Compliance

 __  __ _            _    ___
|  \/  (_)_ __   ___| |  |_ _|
| |\/| | | '_ \ / _ \ |   | |
| |  | | | | | |  __/ |   | |
|_|  |_|_|_| |_|\___|_|  |___|

Mine AI is an open-source toolkit and command-line workspace for advancing AI-enabled payment-security, fraud-detection, AML/CFT compliance, and risk-control infrastructure for U.S.-accessible digital-asset and blockchain-based payment systems — through graph-temporal deep-learning models, production-grade transaction reconciliation and risk controls, and LLM-powered interpretable compliance reasoning.

It is structured as three independently usable modules that together cover the detection layer, the operational risk-control layer, and the interpretability/compliance-reasoning layer of a modern payment-security stack.

This is research and reference-implementation infrastructure, not a production fraud-detection product and not financial, legal, or compliance advice.

Payment fraud and digital-asset-enabled financial crime are documented U.S. national-scale risk channels.

• FBI IC3 2025 Annual Report reports approximately $11.367 billion in cryptocurrency-related losses and approximately $8.649 billion in investment-fraud losses in 2025.
• FinCEN National AML/CFT Priorities identify cybercrime (with virtual-currency considerations) and fraud as significant AML/CFT threats.
• OFAC has designated cryptocurrency mixing and cross-chain laundering infrastructure as sanctions targets, including Tornado Cash and Blender.
• Executive Order 14178 (January 23, 2025) directs the development of a federal digital-asset regulatory framework covering market structure, oversight, consumer protection, and risk management.
• Anti-Money Laundering Act of 2020 (Public Law 116-283, Title LXIV) modernized BSA-driven AML enforcement.
• OCC / Federal Reserve / FDIC 2026 revised interagency model-risk-management guidance emphasizes model development, validation, monitoring, governance, controls, and third-party model review — explicitly acknowledging further work on banks' use of generative and agentic AI.
• 2024 Critical and Emerging Technologies List identifies artificial intelligence, large language models, distributed ledger technologies, digital assets, and digital payment technologies as priority areas.

Industry detection systems still rely heavily on rule-based heuristics — address blacklists, transaction-velocity thresholds, simple graph-search, and manual investigation workflows. These approaches struggle with the severe class imbalance characteristic of fraud datasets, the graph-structured and temporal nature of payment networks, and the interpretability requirements of regulated compliance review. Mine AI exists to make better methods reproducible and inspectable in the open.

Mine AI is organized as three modules that map to the detection, operations, and compliance-reasoning layers of a payment-security stack:

+-----------------------------------------------------------------+
|  Module 3 — LLM-Powered Interpretable Compliance Reasoning      |
|    RAG over BSA / FinCEN / OFAC / EO 14178 corpus               |
|    Interpretable chain-of-thought traces for review workflows   |
+-----------------------------------------------------------------+
                              ^
                              |
+-----------------------------------------------------------------+
|  Module 1 — Graph-Temporal Fraud Detection Models               |
|    CNN-LSTM hybrid (BDICN 2026)                                 |
|    CSSA cross-modal alignment (LLM + graph contrastive)         |
|    FinSCRA multi-chain interpretable reasoning                  |
+-----------------------------------------------------------------+
                              ^
                              |
+-----------------------------------------------------------------+
|  Module 2 — Production Risk-Control & Reconciliation Patterns   |
|    On-chain / off-chain reconciliation reference                |
|    Health-factor monitoring + automated rebalancing patterns    |
|    Adversarial-execution and oracle-manipulation defense notes  |
+-----------------------------------------------------------------+

If detection models cannot trust the underlying ledger state (Module 2), their outputs are unreliable; if their outputs cannot be explained (Module 3), they are difficult to operate in regulated workflows. The three modules are designed to be used together but can be adopted independently.

Reference implementations of three published or peer-citable architectures for payment-fraud and AML/CFT detection on blockchain payment data. The PyTorch / Python source for each architecture lives in models/ — see models/README.md for the directory map and per-model faithfulness notes.

• Paper: Detection of Blockchain Online Payment Fraud Via CNN-LSTM, BDICN 2026 (5th International Conference on Big Data, Information and Computer Network), ACM Digital Library, DOI 10.1145/3801228.3801323.
• The CNN component captures local structural patterns in graph-derived transaction features; the LSTM component captures temporal dependencies in transaction sequences. The hybrid is engineered to address the imbalanced-dataset problem through targeted sampling and loss-weighting strategies described in the paper.
• Methodologically distinct from standalone GNN classifiers (which lack temporal sensitivity) and standalone recurrent classifiers (which lack graph-structural inductive bias).

• Paper: Zhao, Yuan, Wang, Shen, Huang. CSSA: A Cross-modal Semantic-Structural Alignment Framework via LLMs and Graph Contrastive Learning for Fraud Detection of Online Payment. CNML 2026 (IEEE), DOI 10.1109/CNML68938.2026.11452378.
• Integrates LLM semantic representations with graph-contrastive structural representations through a unified contrastive objective. Addresses the gap left by separate semantic (NLP-on-memos) and structural (GNN-on-graph) models that lack aligned cross-modal representation.

• Paper: Pan, Chen, He, Yuan, Wang, Zhang. FinSCRA: An LLM-Powered Multi-Chain Reasoning Framework for Interpretable Node Classification on Text-Attributed Graphs. ICCECE 2026 (IEEE), DOI 10.1109/ICCECE69169.2026.11399797.
• Introduces explicit chain-of-thought reasoning over heterogeneous on-chain entities. Interpretability-by-construction, designed to be operable in BSA / model-governance workflows where black-box graph classifiers struggle.

# Inspect the model card / paper / feature snapshot via the CLI
mineai detect --model cnn-lstm --input ./samples/btc-fraud-sample.json

# Or run the actual reference PyTorch implementation (CPU, seconds)
cd models/cnn_lstm && pip install -r requirements.txt && python demo.py
cd models/cssa     && pip install -r requirements.txt && python demo.py
cd models/finscra  && python demo.py   # needs OPENAI_API_KEY

A documentation and reference-pattern library for the operational substrate that fraud-detection models depend on. Sourced from publicly observable DeFi-engineering patterns; does not republish any proprietary protocol code.

Covers:

• On-chain / off-chain reconciliation — connecting smart-contract state with backend accounting so that downstream detectors have reliable ground truth.
• Health-factor monitoring and automated rebalancing — patterns for collateralized lending and BTCFi vault environments.
• Adversarial-execution defense — engineering notes on MEV-aware execution paths.
• Oracle-manipulation defense — design patterns for price-feed validation and circuit-breakers.
• Liquidation-cascade resilience — operational notes informed by publicly reported cascade events.

# Generate a reconciliation-pattern scaffold
mineai pattern reconciliation --chain sui --output ./my-recon-app/

# Browse the pattern catalog
mineai pattern list

A retrieval-augmented reasoning workbench over public U.S. payment-security and AML/CFT corpora. Designed to produce traceable, citation-grounded reasoning rather than opaque classifications.

Corpus (open sources only):

• Bank Secrecy Act (31 U.S.C. § 5311 et seq.)
• FinCEN National AML/CFT Priorities and ransomware-related BSA materials
• OFAC sanctions designations (including Tornado Cash, Blender)
• Treasury 2024 National Strategy for Combating Terrorist and Other Illicit Financing
• Anti-Money Laundering Act of 2020
• OCC / Federal Reserve / FDIC 2026 revised model-risk-management guidance
• Executive Order 14178
• 2024 Critical and Emerging Technologies List

# Ask an interpretable, citation-grounded compliance-research question
mineai reason "What BSA documentation expectations apply to a U.S. MSB
               offering on-chain swap aggregation across multiple chains?"

# Inspect the underlying reasoning trace
mineai reason --trace last

This module produces research-grade reasoning aids. It does not file SARs, render legal opinions, or replace qualified compliance counsel.

Node.js 20 or higher.

npm install -g @mine-ai/mine-ai
mineai --version

Mine AI is model-agnostic and does not ship with a hosted inference service. Point it at any OpenAI-compatible endpoint you already have access to:

export OPENAI_API_KEY="..."
export OPENAI_BASE_URL="https://your-endpoint/v1"
export OPENAI_MODEL="your-model-id"

The roadmap maps to the technical agenda this project exists to support.

• Publish a public benchmark dataset of labeled digital-asset payment-fraud and AML/CFT cases, sourced from public on-chain data, OFAC sanctions designations, and open fraud reports.
• Release three production-pattern technical write-ups covering reconciliation, health-factor monitoring, and BTCFi-vault smart-contract integration.
• Publish a benchmark-methodology technical report; target dissemination venue includes the NeurIPS Datasets & Benchmarks Track.

• Extend the CNN-LSTM, CSSA, and FinSCRA architectures with adversarial-robustness evaluation, ablation studies, and replication on the Phase-I benchmark.
• Target peer-review venues: AAAI, IJCAI, IEEE Symposium on Security and Privacy, Financial Cryptography and Data Security, KDD, ACL Industry Track, EMNLP, NeurIPS, ICML.
• Add adversarial-execution defense, oracle-manipulation defense, and liquidation-cascade-resilience benchmarking to Module 2.

• Public workshops, open-source reference releases, and standards or public-comment contributions where appropriate.
• Outreach to compliance, RegTech, digital-asset risk, and model-governance practitioner communities.

• Not a yield-farming, staking, or mining tool. Mine AI uses digital-asset systems as a payment-data environment for studying fraud and risk, not as an investment opportunity.
• Not a node-deployment helper. Earlier iterations of this repository focused on validator and node operations; that scope has been retired.
• Not financial, legal, tax, or compliance advice.
• Not a production SAR-filing or sanctions-screening system. Outputs are research aids that require qualified human review before any regulated use.
• Not a hosted service. There is no paid SaaS tier, no required signup, and no proprietary API.

• Yuan, Lin, Wu, Chang. Detection of Blockchain Online Payment Fraud Via CNN-LSTM. BDICN 2026, ACM Digital Library. DOI 10.1145/3801228.3801323.
• Zhao, Yuan, Wang, Shen, Huang. CSSA: A Cross-modal Semantic-Structural Alignment Framework via LLMs and Graph Contrastive Learning for Fraud Detection of Online Payment. CNML 2026 (IEEE). DOI 10.1109/CNML68938.2026.11452378.
• Pan, Chen, He, Yuan, Wang, Zhang. FinSCRA: An LLM-Powered Multi-Chain Reasoning Framework for Interpretable Node Classification on Text-Attributed Graphs. ICCECE 2026 (IEEE). DOI 10.1109/ICCECE69169.2026.11399797.

• FBI IC3 Annual Reports (2024, 2025).
• FinCEN National AML/CFT Priorities.
• OFAC sanctions designations (Tornado Cash; Blender).
• Treasury, 2024 National Strategy for Combating Terrorist and Other Illicit Financing.
• Anti-Money Laundering Act of 2020 (Public Law 116-283, Title LXIV).
• Executive Order 14178 (January 23, 2025), Strengthening American Leadership in Digital Financial Technology.
• OCC / Federal Reserve / FDIC, 2026 revised interagency model-risk-management guidance.
• White House / NSTC, 2024 Critical and Emerging Technologies List.
• Bank Secrecy Act, 31 U.S.C. § 5311 et seq.

Contributions are welcome — particularly benchmark contributions, reference-implementation reproductions, and pattern write-ups grounded in publicly observable systems. See CONTRIBUTING.md.

LICENSE