Mine AI: Reference Implementations for AI-Powered Blockchain Fraud Detection

What Mine AI Addresses in Payment Security

Payment fraud and AML/CFT detection in digital-asset systems still rely heavily on rule-based heuristics: address blacklists, transaction-velocity thresholds, simple graph search, and manual investigation workflows. The Mine AI README cites documented shortcomings of these approaches: they struggle with the severe class imbalance typical 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. The README identifies three published regulatory and law-enforcement contexts motivating the work: FBI IC3 2025 reporting on cryptocurrency-related losses, FinCEN National AML/CFT Priorities identifying cybercrime as a significant threat, and the 2024 Critical and Emerging Technologies List naming AI, LLMs, distributed ledger technologies, and digital payment technologies as priority areas.

The toolkit is structured as three independently usable modules that map to the detection layer, the operational risk-control layer, and the interpretability and compliance-reasoning layer of a payment-security stack.

Module 1: Three Graph-Temporal Fraud Detection Model Architectures

Module 1 provides reference implementations of three published architectures for payment fraud and AML/CFT detection on blockchain payment data. The source lives in models/.

The first is a CNN-LSTM hybrid for blockchain payment fraud, based on the paper 'Detection of Blockchain Online Payment Fraud Via CNN-LSTM' published at BDICN 2026 and indexed in the ACM Digital Library under DOI 10.1145/3801228.3801323. The CNN component captures local structural patterns in graph-derived transaction features; the LSTM captures temporal dependencies in transaction sequences. The architecture is designed to handle imbalanced datasets through targeted sampling and loss-weighting strategies described in the paper.

The second is CSSA (Cross-Modal Semantic-Structural Alignment), published at CNML 2026 (IEEE) under DOI 10.1109/CNML68938.2026.11452378. CSSA integrates LLM semantic representations with graph-contrastive structural representations through a unified contrastive objective, addressing the gap between separate NLP and GNN models that lack aligned cross-modal representation.

The third is FinSCRA (LLM-Powered Multi-Chain Reasoning), published at ICCECE 2026 (IEEE) under DOI 10.1109/ICCECE69169.2026.11399797. FinSCRA introduces explicit chain-of-thought reasoning over heterogeneous on-chain entities, designed to produce interpretable outputs compatible with BSA and model-governance workflows where black-box classifiers are difficult to operate.

Running the Model Demos

Each model ships a demo.py file and a requirements.txt in its own subdirectory under models/. To inspect a model's documentation and run it:

mineai detect --model cnn-lstm --input ./samples/btc-fraud-sample.json

This CLI path uses the mineai command from the npm package. To run the PyTorch implementation directly:

cd models/cnn_lstm && pip install -r requirements.txt && python demo.py

For the CSSA model:

cd models/cssa && pip install -r requirements.txt && python demo.py

For FinSCRA, which uses an LLM for reasoning:

cd models/finscra && python demo.py

FinSCRA requires an OPENAI_API_KEY environment variable to be set before running. The repository includes a btc-fraud-sample.json in samples/ for testing the CLI path. The package.json lists Node.js >= 20 as the engine requirement for the CLI wrapper.

Module 2 and Module 3: Risk Control and Compliance Reasoning

Module 2 covers production risk-control and reconciliation patterns. The README describes this as a documentation and reference-pattern module, covering operational concerns around the detection outputs: transaction reconciliation, risk controls, and the reliability of the underlying ledger state that detection models depend on. The README notes that if detection models cannot trust the underlying ledger state, their outputs are unreliable regardless of model quality.

Module 3 covers LLM-powered interpretable compliance reasoning. The architecture described in the README is a RAG (Retrieval-Augmented Generation) system built over a corpus of BSA, FinCEN, OFAC, and Executive Order 14178 documents. The LLM produces chain-of-thought traces intended for use in compliance review workflows where a plain classification output is insufficient.

The three modules are designed to work together but can each be adopted independently. A team that only needs the fraud detection models can use Module 1 without Module 2 or 3.

Research Context and Regulatory Framing

The README cites several specific regulatory and enforcement documents as context for the work: the Anti-Money Laundering Act of 2020 (Public Law 116-283, Title LXIV) as the modernizing BSA-driven AML enforcement framework; OFAC designations of cryptocurrency mixing and cross-chain laundering infrastructure as sanctions targets; and the OCC/Federal Reserve/FDIC 2026 revised interagency model-risk-management guidance, which the README notes explicitly acknowledges banks' use of generative and agentic AI.

The README is careful to frame the repository as research and reference-implementation infrastructure, not financial, legal, or compliance advice and not a production fraud-detection product. This framing is relevant for anyone considering the repository for regulated use: the model implementations reference published papers but have not been independently validated against production payment data.

Limitations and Comparison to Production Fraud Systems

Mine AI is explicitly a reference implementation. The models in Module 1 reproduce architectures from academic papers and include demo scripts that run in seconds on CPU against sample data. They are not trained on production-scale payment datasets and are not calibrated for any specific institution's fraud distribution.

FinSCRA's dependency on an OpenAI API key means it requires a paid external service and is subject to that service's availability and pricing. Organizations operating in restricted environments may not be able to use it without a local model substitution.

An alternative in the production fraud detection space is Graph Neural Network-based systems offered by commercial vendors, which are trained on proprietary transaction data at scale and validated against real fraud rates. Those systems are not inspectable or reproducible. Mine AI's distinction is that every architecture is backed by a published paper with a DOI, the source code is MIT-licensed and auditable, and the implementations can be reproduced on public or synthetic data without a vendor relationship.

The last push to this repository was on 2026-05-13. The project is MIT licensed.

Editorial conclusion

Mine AI is appropriate for researchers and compliance engineers who want inspectable reference implementations of published fraud-detection architectures and documented risk-control patterns, not a ready-to-deploy product. The repository explicitly frames itself as research infrastructure, not production fraud detection software. Before using any model in a workflow, read the module README in models/ and understand the paper's stated assumptions, since dataset characteristics and class imbalance ratios differ across deployments. The last push was on 2026-05-13.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. mine-ai-xyz/mine-ai on GitHub
4. README