Blockchain Fraud Detection

Fraud Detection in Blockchain Networks

Research Lead: Prof. Yuanyuan Zhang & Prof. Stephen Chan Primary Focus: Graph-based methods for detecting fraudulent activities in blockchain transactions

Research Overview

Blockchain networks generate massive transaction graphs where fraudulent actors leave detectable patterns. Our research develops advanced graph neural network methods to identify fraud in cryptocurrency ecosystems.

Key Research Areas

Ethereum Transaction Analysis

Detecting fraudulent smart contracts and Ponzi schemes in the Ethereum network.

Multi-Layer Graph Learning

Combining different graph views (transaction, contract, address) for improved detection.

Topology-Aware Contrastive Learning

Novel self-supervised methods that leverage blockchain graph structure.

The Blockchain Fraud Problem

Types of Blockchain Fraud

Type Description Detection Challenge
Ponzi Schemes Pyramid-style smart contracts Complex fund flows
Phishing Fake addresses/contracts Address similarity
Exit Scams Rug pulls on DeFi Timing prediction
Wash Trading Artificial volume Transaction pattern
Money Laundering Fund obfuscation Multi-hop analysis

Scale of the Problem

  • Over $14 billion lost to crypto fraud in 2022
  • Ethereum processes 1M+ transactions daily
  • Manual review impossible at blockchain scale
  • Fraudsters constantly evolve tactics

Methodology: Multilayer Topology-Aware GCL

Overview

Our published research introduces Multilayer Topology-Aware Graph Contrastive Learning (MT-GCL) for fraud detection.

Architecture

Input: Ethereum Transaction Graph
       |
   [Layer 1: Transaction Subgraph]
   [Layer 2: Contract Call Graph]
   [Layer 3: Token Transfer Graph]
       |
   [Topology-Aware Augmentation]
       |
   [Contrastive Learning Encoder]
       |
   [Multi-Layer Fusion]
       |
Output: Fraud Probability per Address

Key Innovations

  1. Multi-Layer Representation: Captures different relationship types simultaneously
  2. Topology-Aware Augmentation: Preserves important graph structures during learning
  3. Self-Supervised Pretraining: Reduces need for labeled fraud examples
  4. Scalable Architecture: Handles millions of addresses

Experimental Results

Dataset: Ethereum Transaction Network

Statistic Value
Addresses 2.9M
Transactions 13.5M
Smart Contracts 58K
Labeled Frauds 4,328
Time Period 2015-2021

Performance Comparison

Method AUC-ROC Precision Recall F1
Random Forest 0.721 0.189 0.312 0.235
XGBoost 0.756 0.214 0.356 0.267
GCN 0.812 0.287 0.423 0.342
GAT 0.834 0.312 0.456 0.371
Graph-SAGE 0.841 0.324 0.467 0.383
MT-GCL (Ours) 0.891 0.398 0.534 0.456

Ablation Study

Variant AUC-ROC Change
MT-GCL (Full) 0.891 -
Single Layer 0.856 -3.9%
No Contrastive 0.847 -4.9%
Random Augmentation 0.863 -3.1%
No Pretraining 0.858 -3.7%

Detection Examples

Case Study: Ponzi Scheme Detection

Identified Pattern

Star topology with central contract receiving funds from many addresses and distributing to few early investors.

Case Study: Phishing Attack Detection

Identified Pattern

Address clusters with similar naming patterns targeting legitimate DeFi protocols.

Publications

Multilayer Topology-Aware Graph Contrastive Learning for Fraud Detection in the Ethereum Transaction Network

Chen, Y., Zhang, Y., Chan, S., Chu, J., & Lord, N.

Journal of the Royal Statistical Society Series A - In Press

DOI | PDF

Practical Applications

For Exchanges

Application Benefit
Deposit Screening Block fraudulent funds
Account Monitoring Detect suspicious behavior
Compliance Meet regulatory requirements

For Regulators

Application Benefit
Market Surveillance Identify market manipulation
Investigation Support Trace illicit funds
Risk Assessment Evaluate platform risks

For DeFi Protocols

Application Benefit
Contract Auditing Identify risky contracts
User Protection Warn about suspicious addresses
Insurance Risk-based pricing

Code & Resources

Resource Status
Paper Code Available on GitHub
Trained Models Available on request
Dataset Ethereum public data
Documentation Paper supplementary

Future Directions

  1. Real-Time Detection: Streaming graph analysis for instant fraud alerts
  2. Cross-Chain Analysis: Detecting fraud across multiple blockchains
  3. Privacy-Preserving: Detection without revealing sensitive patterns
  4. Explainable AI: Human-interpretable fraud explanations

(c) Joerg Osterrieder 2025