---
license: mit
task_categories:
- tabular-classification
language:
- en
tags:
- ethereum
- fraud-detection
- blockchain
- multimodal
- xgboost
- shap
- explainable-ai
pretty_name: FuseChain Ethereum Fraud Detection Model
---

# FuseChain: Ethereum Fraud Detection via Multimodal Signal Fusion

## Model Summary

FuseChain is a multimodal supervised classification model for detecting fraudulent Ethereum Externally Owned Accounts (EOAs). It integrates on-chain transaction features with off-chain contextual signals from market data, Reddit, and Twitter to classify Ethereum addresses as scam or normal.

The model is an **XGBoost classifier** trained on a novel address-level dataset of **35,272 Ethereum EOAs**, achieving an **F1-score of 82.5%** and an **AUC of 96.1%** on a stratified held-out test set — representing a **14.7 point F1 improvement** over an on-chain only baseline.

---

## Model Details

| Property | Details |
|---|---|
| Model Type | XGBoost Classifier |
| Task | Binary Classification (Scam / Normal) |
| Input | 31 address-level multimodal features |
| Output | Fraud probability score (0 to 1) |
| Classification Threshold | 0.5 |
| Explainability | TreeSHAP (per-prediction feature attribution) |
| Training Framework | XGBoost 2.x, Scikit-learn |
| Language | Python 3.10+ |

---

## Performance

### Test Set Results (Stratified 80/20 Split)

| Metric | Normal | Scam | Overall |
|---|---|---|---|
| Precision | 0.96 | 0.89 | 0.95 |
| Recall | 0.98 | 0.77 | 0.95 |
| F1-Score | 0.97 | 0.83 | 0.95 |
| AUC-ROC | - | - | 0.961 |
| Accuracy | - | - | 95% |

### Ablation Study Results

| Configuration | Features | F1 | AUC |
|---|---|---|---|
| On-Chain Only | 14 | 0.678 | 0.919 |
| On-Chain + Market | 19 | 0.721 | 0.936 |
| On-Chain + Market + Reddit | 22 | 0.802 | 0.955 |
| On-Chain + Market + Twitter | 28 | 0.825 | 0.962 |
| On-Chain + Market + Reddit + Twitter | 31 | 0.825 | 0.961 |

---

## Feature Set

The model was trained on 31 features across four modalities:

| Modality | Features | Examples |
|---|---|---|
| On-Chain | 14 | `eth_net_flow_max`, `eth_recv_mean`, `burst_max_tx_5m_mean`, `active_days` |
| Twitter | 9 | `twitter_avg_retweets_mean`, `twitter_avg_positive_mean`, `twitter_fraud_mention_ratio_mean` |
| Market | 5 | `market_intraday_volatility_mean`, `market_daily_return_mean` |
| Reddit | 3 | `reddit_total_fraud_mentions_mean`, `reddit_avg_sentiment_mean` |

For the full feature schema refer to `address_features_metadata.json` in this repository.

---

### Global Modality Contribution (SHAP)

| Modality | Contribution |
|---|---|
| On-Chain | 58.6% |
| Twitter | 25.8% |
| Reddit | 10.8% |
| Market | 4.8% |

### Most Discriminative Features per Modality

| Modality | Top Feature |
|---|---|
| On-Chain | `eth_net_flow_max` |
| Twitter | `twitter_avg_retweets_mean` |
| Reddit | `reddit_total_fraud_mentions_mean` |
| Market | `market_intraday_volatility_mean` |

---

## Model Hyperparameters

| Parameter | Value |
|---|---|
| n_estimators | 200 |
| max_depth | 5 |
| learning_rate | 0.05 |
| min_child_weight | 3 |
| subsample | 0.8 |
| colsample_bytree | 0.8 |
| Classification Threshold | 0.5 |

---

## Dataset

The FuseChain dataset used to train this model is publicly available on Hugging Face:

[FuseChain Multimodal Ethereum Fraud Dataset](https://huggingface.co/datasets/Nileshka/fusechain-data)


## Citation

If you use this model or the FuseChain framework in your research, please cite:
```bibtex
@misc{fusechain2026,
  title={FuseChain: Ethereum Fraud Detection via Multimodal Signal Fusion},
  author={Fernando, Nileshka},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/Nileshka/fusechain-data}}
}
```

---

## Related Resources

- **Dataset:** [FuseChain Multimodal Ethereum Fraud Dataset](https://huggingface.co/datasets/Nileshka/fusechain-data)
- **Code Repository:** [FuseChain GitHub](https://github.com/NileshFdo/FuseChain-FYP)