Detecting Sybil Addresses in Blockchain Airdrops: A Subgraph-based Feature Propagation and Fusion Approach

Qiangqiang Liu, Qian Huang, Frank Fan, Haishan Wu, Xueyan Tang

Abstract

Airdrops are a cornerstone of Web3 marketing, yet they are frequently undermined by Sybil attacks—where single entities control numerous accounts to monopolize rewards. Traditional detection methods often struggle with scalability or the evolving sophistication of attackers. We present Subgraph-based lightGBM, a supervised learning framework that identifies Sybil addresses by analyzing localized transaction subgraphs and temporal lifecycle patterns. By propagating features across a two-layer neighborhood, our model achieves a precision of 94.28% and an AUC of 0.9806 on real-world data from the Binance Account Bound (BAB) token airdrop, significantly outperforming existing benchmarks.

Introduction

In decentralized ecosystems, airdrops aim to distribute tokens to genuine community members. However, "airdrop farmers" deploy automated Sybil attacks to drain these resources. Detecting these actors is challenging because full blockchain transaction graphs are computationally prohibitive to analyze, and unsupervised clustering methods often yield high false-positive rates or require extensive manual tuning. This research proposes a scalable, automated approach focusing on local subgraph structures and "just-in-time" behavioral signatures.

Methodology

The proposed system utilizes a multi-stage pipeline centered on localized graph analysis and gradient boosting.

Two-Layer Subgraph Construction

To maintain computational efficiency, the model constructs a local "neighborhood" for each target address rather than processing the entire global graph. This involves capturing transactions two steps upstream (funding sources) and two steps downstream (fund dispersal).

Two-layer subgraph construction and feature propagation process.

Feature Extraction and Propagation

Three primary feature categories are extracted:

  1. Temporal (Lifecycle): Metrics such as the interval between initial funding ("gas"), first transaction, and airdrop participation. Sybils typically exhibit highly synchronized, "tight" temporal clustering.
  2. Amount-based: Statistical summaries of account balances and transfer volumes.
  3. Network Structure: Topological patterns of transactions.

A key innovation is Feature Propagation: attributes from neighboring nodes are fused into the target address's profile. For instance, if a target is funded by a hub that also funded hundreds of other accounts, this structural context is explicitly captured.

Experimental Results

The model was evaluated using a dataset from the 2022 Binance Account Bound (BAB) token event, comprising 193,701 addresses and over 58 million transactions. Labels were enriched using Arkham Intelligence to filter out institutional wallets and smart contracts.

Identifying Attack Topologies

Analysis reveals that legitimate users typically possess "noisy" transaction histories, whereas Sybil accounts follow a rigid, linear lifecycle. Attackers frequently employ specific topologies to manage their fleets efficiently.

Common Sybil attack topologies: Star, Tree, and Chain patterns.

"Star" patterns, where a central hub distributes funds to many "spokes," are particularly common as they minimize transaction fees while maintaining control. Our subgraph approach is specifically tuned to detect these structural red flags, often identifying Sybils with 95% to 100% probability in these configurations.

Performance Comparison

The Subgraph-based lightGBM model demonstrates superior performance across all key metrics compared to standard machine learning baselines and industry-standard tools like Trusta.

Method Precision Recall F1-Score AUC
Subgraph-based lightGBM 0.9428 0.9182 0.9303 0.9806
Clustering-based Trusta 0.7962 0.8159 0.8059 0.8642
lightGBM 0.7821 0.7150 0.7470 0.8484
Decision Tree (DT) 0.7289 0.7572 0.7428 0.7373
SVM 0.6203 0.0329 0.0624 0.5599

Conclusion

By shifting from global graph analysis to localized subgraph feature propagation, we provide a high-precision, scalable solution for Sybil detection. This supervised approach minimizes false positives, ensuring that legitimate users are not unfairly excluded from airdrop rewards. Future work will explore Graph Neural Networks (GNNs) to further enhance the interpretability of these structural detections.

Summary updated

BibTeX
@inproceedings{liu2025detecting,
  author = {Qiangqiang Liu and Qian Huang and Frank Fan and Haishan Wu and Xueyan Tang},
  title = {{Detecting Sybil Addresses in Blockchain Airdrops: A Subgraph-based Feature Propagation and Fusion Approach}},
  booktitle = {ICBC},
  year = {2025},
  doi = {10.1109/icbc64466.2025.11185061},
}
Figure