AML detection in blockchain systems

A comprehensive final-year research project developing an Anti-Money Laundering (AML) detection mechanism for decentralized blockchain systems. This project implements multiple ML/AI approaches to identify known money laundering patterns and validates detection effectiveness on real blockchain networks.

Research Poster

Below is the visual research poster presented for this final-year project, which summarizes the methodology, key findings, and contributions of the AML detection system.

Poster Overview

The research poster visually presents the comprehensive four-phase methodology used to develop and validate the AML detection system. It highlights:

• Research Framework: The systematic approach from testnet development through mainnet validation to advanced AI techniques
• Technical Architecture: Visual representation of the ML/AI pipeline including feature extraction, model training, and pattern detection
• Key Results: Performance metrics and detection rates across different blockchain networks
• Historical Case Studies: Visual analysis of real-world money laundering cases including Bitfinex, PlusToken, and Ronin Bridge
• Contributions: The project's impact on blockchain AML research and practical applications

The poster demonstrates the project's scope, from initial pattern-based detection algorithms to advanced Graph Neural Networks and LSTM implementations, providing a comprehensive overview of the research journey and outcomes.

Project Overview

This experimental study investigates AML techniques in blockchain systems and develops a pattern-based detection mechanism. The system analyses transaction patterns across Ethereum and TRON (TRC-20) networks to identify money laundering activities including fanout, layering, and mixing techniques.

Research Objectives

The project aims to:

• Study existing AML techniques for blockchain systems
• Implement detection algorithms based on known money laundering patterns
• Evaluate ML/AI approaches (Random Forest, Neural Networks, GNN, LSTM, Autoencoders)
• Validate detection mechanisms on real blockchain networks
• Analyse historical money laundering cases (Bitfinex 2016, PlusToken 2019, Ronin Bridge 2022)

Project Goals

What We Want to Build:

An advanced AML detection system that can:

• Identify money laundering patterns in blockchain transactions in real-time
• Detect known laundering techniques: fanout, layering, mixing, and money mule networks
• Provide risk scoring for addresses and transactions
• Analyse historical cases to validate detection effectiveness
• Compare performance across different blockchain networks (Ethereum and TRON)

Target Outcomes:

• Production-ready detection system with documented methodology
• Comparative analysis of ML/AI approaches for blockchain AML
• Validation against real-world historical cases
• Framework applicable to multiple blockchain networks

Methodology: Four-Phase Approach

The project follows a systematic four-phase methodology, with Phase 1 completed, Phase 2 in progress, and Phases 3 and 4 planned for completion.

Phase 1: Testnet AI Development

Status: Completed

Objective: Build and test ML/AI detection algorithms on Ethereum Goerli testnet

What Was Done:

• Developed pattern-based detection following known money laundering techniques
• Extracted 50+ features from transaction patterns including:
  • Transaction frequency and velocity
  • Amount patterns and distributions
  • Network topology metrics
  • Temporal patterns
  • Address clustering characteristics
• Implemented multiple ML models:
  • Random Forest: Baseline classification for pattern detection
  • Neural Networks: Deep learning approach for complex pattern recognition
  • Ensemble Systems: Combining multiple models for improved accuracy

Results Achieved:

• Established baseline detection performance
• Validated feature extraction methodology
• Confirmed pattern-based approach effectiveness on testnet
• Created foundation for mainnet deployment

Phase 2: Mainnet Validation

Status: In progress

Objective: Validate detection mechanism on Ethereum mainnet with real transactions

What Was Done:

• Deployed detection system on Ethereum mainnet with controlled amounts ($50–100)
• Compared pattern-based AML system against commercial tools
• Measured real-world detection rates for known laundering patterns
• Evaluated false positive rates and system performance

Key Findings:

• Pattern-based approach identified known laundering techniques effectively
• Detection rates comparable to commercial solutions
• System demonstrated ability to identify fanout and layering patterns
• Low false positive rate achieved with proper tuning

Phase 3: Advanced AI Techniques

Status: In progress

Objective: Implement cutting-edge AML detection methods

Planned Implementation:

1. Graph Neural Networks (GNN)

   • Will analyse transaction graph structures
   • Will identify money mule networks and complex laundering schemes
   • Will detect suspicious network topologies
   • Will map relationships between addresses and transactions

2. LSTM (Long Short-Term Memory) Networks

   • Will perform temporal pattern detection
   • Will conduct sequence analysis of transaction flows
   • Will enable time-series prediction for suspicious activity
   • Will capture long-term dependencies in transaction sequences

3. Autoencoders

   • Will enable anomaly detection through reconstruction error
   • Will utilise unsupervised learning for unknown patterns
   • Will identify novel laundering techniques
   • Will detect deviations from normal transaction behaviour

Expected Impact: Advanced techniques will improve detection of complex, multi-stage laundering operations that traditional ML approaches may miss.

Phase 4: Historical Case Analysis

Status: Planned

Objective: Test detection mechanism against real-world historical cases

Planned Case Studies:

1. Bitfinex Hack 2016

   • Will analyse $72 million theft and subsequent laundering
   • Will trace transaction flows through blockchain
   • Will evaluate detection system's ability to identify patterns retrospectively
   • Will validate system against known laundering paths

2. PlusToken Scam 2019

   • Will analyse $2.9 billion cryptocurrency scam
   • Will conduct complex laundering network analysis
   • Will perform multi-blockchain transaction tracking
   • Will test system on large-scale laundering operations

3. Ronin Bridge Hack 2022

   • Will analyse $625 million theft
   • Will perform rapid laundering pattern analysis
   • Will evaluate real-time detection capability
   • Will test system response to fast-moving laundering schemes

Expected Results: Detection system will demonstrate practical applicability by successfully identifying laundering patterns in historical cases.

Technical Implementation

Technologies Used

• Python: Core development language
• SQL: Transaction data storage and querying
• Machine Learning Libraries: Scikit-learn, TensorFlow, PyTorch
• Blockchain Analysis: Web3.py, TronPy
• Data Visualisation: Matplotlib, Plotly, Splunk integration
• Graph Analysis: NetworkX, DGL (Deep Graph Library)

Key Features

• Pattern Recognition: Identifies fanout, layering, mixing, and other known techniques
• Real-time Monitoring: Continuous analysis of blockchain transactions
• Risk Scoring: Generates risk scores for addresses and transactions
• Network Analysis: Maps money mule networks and laundering structures
• Historical Analysis: Retrospective analysis of past laundering cases

Research Questions

The project addresses the following research questions:

1. What AML techniques are most effective for blockchain systems?

   • Comparing traditional ML, deep learning, GNN, and autoencoder approaches
   • Evaluating performance across different pattern types

2. How effective is pattern-based detection compared to anomaly detection?

   • Measuring detection rates and false positive rates
   • Comparing against commercial solutions

3. How do detection rates vary across different blockchain networks?

   • Ethereum vs TRON (TRC-20) comparison
   • Network-specific feature engineering impact

4. What are the performance characteristics of different ML/AI approaches?

   • Speed, accuracy, interpretability trade-offs
   • Resource requirements and scalability

Current Status

Completed Work:

Phase 1 (Testnet AI Development) has been successfully completed. This phase established the foundation of the detection system, validated the pattern-based approach, and demonstrated effectiveness on testnet environments.

In Progress:

Phase 2 (Mainnet Validation) is currently in progress. This phase involves deploying the detection system on Ethereum mainnet with real transactions and comparing performance against commercial tools.

In Progress:

Phase 3 (Advanced AI Techniques) is currently in progress. This phase involves implementing Graph Neural Networks, LSTM networks, and Autoencoders to enhance detection capabilities for complex laundering operations.

Planned:

Phase 4 (Historical Case Analysis) is planned for future implementation. This phase will validate the system against real-world historical cases including the Bitfinex hack, PlusToken scam, and Ronin Bridge hack.

Contributions

This project contributes to the field of blockchain AML by:

• Comprehensive Evaluation: First study comparing multiple ML/AI approaches for blockchain AML
• Real-world Validation: Testing on actual mainnet transactions and historical cases
• Pattern-based Methodology: Systematic approach to detecting known laundering techniques
• Cross-network Analysis: Comparison of effectiveness across different blockchain networks
• Practical Framework: Deployable detection system with documented limitations and best practices

Relevance to Fraud Management

This project directly demonstrates skills relevant to fraud prevention and financial crime analysis:

• Fraud Pattern Analysis: Deep understanding of money laundering techniques
• Data Analysis: SQL and Python proficiency for transaction analysis
• Risk Assessment: Developing metrics and scoring systems
• Problem-solving: Breaking down complex fraud detection challenges
• Technical Implementation: Building practical detection systems
• Research Methodology: Systematic approach to validation and testing

Future Work

Beyond the four-phase methodology, future enhancements include:

• Integration with traditional banking AML systems
• Real-time alert generation and case management
• Expansion to additional blockchain networks (Bitcoin, BSC, Polygon)
• Development of explainable AI for regulatory compliance
• Collaboration with financial institutions for validation
• Performance optimisation for large-scale deployment