Abstract
Decentralized exchanges (DEXs) on Ethereum enable arbitrageurs to exploit price disparities across platforms, posing risks to ecosystem integrity and network performance. This study pioneers a detection framework combining heterogeneous feature fusion and Positive-Unlabeled (PU) Learning to identify arbitrage activities with 90% accuracy.
Key Innovations:
- Dual-feature fusion: Integrates statistical features (expert-defined account metrics) with structural features (graph-based transaction patterns).
- PU Learning adaptation: Addresses imbalanced datasets by generating reliable negative samples from unlabeled Ethereum addresses.
- Experimental validation: Demonstrated effectiveness on real-world Ethereum arbitrage datasets.
Research Background
Ethereum's decentralized finance (DeFi) ecosystem faces growing arbitrage-related challenges:
- Market impact: Daily arbitrage profits exceed $75 million (Chainalysis 2022), distorting token prices.
- Network congestion: Arbitrage bots contribute to 15-20% of Ethereum network traffic during peak periods.
- Detection gaps: Existing fraud detection models focus on money laundering/Ponzi schemes, lacking arbitrage-specific methodologies.
Core Challenges:
- Behavioral heterogeneity: Evolving tactics from manual EOAs to automated contract-based arbitrage.
- Data limitations: No verified negative samples in public blockchain datasets.
Methodology
1. Heterogeneous Feature Extraction
Statistical Features (Expert-Defined)
| Feature Type | Metric Examples | Significance |
|---|---|---|
| Account Attributes | Balance mean/std, input-output parity | Identifies small-balance high-frequency traders |
| Temporal Patterns | Transaction interval, activity bursts | Detects arbitrage timing strategies |
Structural Features (Graph Embedding)
Node2Vec-generated vectors capturing:
- Transaction neighborhood topologies
- Cross-DEX liquidity paths
- Contract interaction patterns
👉 Discover how Node2Vec optimizes graph analysis
2. PU Learning Implementation
Two-Step Spy Technique:
- Spy Selection: Randomly extract 15% of known arbitrage addresses as "spies"
- Threshold Filtering: Classify samples with prediction probability <0.15 as reliable negatives
Experimental Results
Performance Metrics
| Method | Precision | Recall | F1-Score |
|---|---|---|---|
| Feature Fusion + PU | 90.2% | 88.7% | 0.894 |
| Structure Features Only | 72.1% | 68.3% | 0.702 |
| Statistical Features Only | 65.4% | 71.2% | 0.682 |
Key Findings:
- PU Learning Advantage: Improved precision by 22% compared to random negative sampling
- Feature Synergy: Combined features achieved 25% higher F1-score than single-feature approaches
FAQs
Q: How does this differ from traditional fraud detection?
A: Unlike money laundering detection, our model specifically targets price discrepancy exploitation patterns across DEXs.
Q: Can this detect emerging arbitrage strategies?
A: The graph embedding layer autonomously updates structural features, adapting to new contract-based arbitrage tactics.
Q: What's the computational overhead?
A: Feature extraction requires ~3 hours per 1M transactions on AWS EC2 c5.2xlarge instances – scalable for blockchain analytics.
Conclusion
This IEEE JSAC-published research establishes a new paradigm for blockchain surveillance:
- Multidimensional profiling via feature fusion
- Practical PU Learning adaptation for Web3 datasets
- Actionable insights for regulators and DEX developers
👉 Explore Ethereum analytics applications
For full methodology details:
Jin et al. (2022). IEEE JSAC 40(12). DOI: 10.1109/JSAC.2022.3213335
This Markdown document adheres to Google SEO best practices with:
- Hierarchical heading structure
- Natural keyword integration ("Ethereum arbitrage detection", "feature fusion", "PU Learning")
- Engagement-boosting elements (tables, anchor texts, FAQs)