Skip to content
Crypto.world

Research

ScamSweeper: Detecting Illegal Accounts in Web3 Scams via Transactions Analysis

Web3 scams imitate legitimate services, then quietly drain user funds and obscure their tracks using mixer contracts or chains of transfers. Existing detection work has focused mostly on phishing…

ScamSweeper: Detecting Illegal Accounts in Web3 Scams via Transactions Analysis
arXiv
2512.15030
Published
2025-12-17
Authors
Xiaoqi Li, Wenkai Li, Zhijie Liu, Meikang Qiu, Zhiquan Liu, Sen Nie, Zongwei Li, Shi Wu, Yuqing Zhang

AI summary

Overview

Research area: Blockchain security — malicious account detection on Ethereum, specifically Web3 scams (cs.CR).

Technical level: Advanced. The paper assumes familiarity with graph neural networks, attention mechanisms, Transformers, and blockchain transaction graphs.

Scope (one sentence): The paper proposes ScamSweeper, a temporal subgraph-sequence learning framework that samples Ethereum transaction networks with a structure-aware, time-aware random walk and learns how malicious accounts' transaction neighborhoods evolve over time to detect Web3 scam accounts.

What This Paper Is About

Web3 scams imitate legitimate services, then quietly drain user funds and obscure their tracks using mixer contracts or chains of transfers. Existing detection work has focused mostly on phishing nodes and de-anonymization, and it typically learns only static graph structure, which struggles on Ethereum's enormous, power-law-distributed, timestamped transaction network. The goal is a detector that captures both the structure of an account's transaction neighborhood and how that neighborhood changes over time, and that can distinguish Web3 scam accounts from both phishing accounts and normal accounts.

Key Contributions

  1. First temporal subgraph-sequence framework for Web3 scam detection on Ethereum. The authors slice an account's transaction graph into subgraphs by time interval and order them as a sequence, so the model learns the dynamic evolution of malicious accounts rather than a single static snapshot.
  2. Structure Temporal Random Walk (STRWalk). An improved sampling method that combines temporal attributes (edges closer in time to the start node get higher sampling weight) with structural expansion (a structure window size w enriches the neighborhood), producing directed subgraphs from a very large network without processing all of it.
  3. A large-scale labeled dataset and attribute analysis. Transactions from the first 18 million block heights on Ethereum, covering Web3 scam, phishing, and normal accounts, with an analysis of differences across nodes, edges, and degree distribution.
  4. Empirical gains over state-of-the-art detectors, plus open-source release. Reported improvements over SIEGE, Ethident, and PDTGA for Web3 scams, and over DGTSG and BERT4ETH for phishing nodes; code and experiment data released on figshare.

Main Findings

  • Web3 scam detection beats three baselines: ScamSweeper outperforms SIEGE, Ethident, and PDTGA, achieving a weighted F1-score improvement of at least 17.29% over a base value of 0.59.
  • Phishing detection also improves: ScamSweeper achieves at least a 17.5% improvement in F1-score over DGTSG and BERT4ETH, starting from a base F1 of 0.80.
  • Malicious and normal accounts differ measurably in graph attributes: The authors report distinctions in node attributes, edge attributes, and degree distribution — particularly in- and out-degree distributions — which motivates building directed graphs rather than undirected ones.
  • Behavioral distinction from phishing: On-chain, traditional phishing accounts simply defraud funds, whereas Web3 scams mimic normal service behavior, steal tokens without user involvement, and obscure traces; this mimicry and track obscuration is the key differentiator the framework targets.
  • Scale motivation: Ethereum transactions reached 2,762 million as of April 2025, which the authors cite as the reason full-graph deep learning is impractical and sampling is required.
  • Component-level ablations (RQ3): The paper poses questions about STRWalk's robustness and whether graph learning or sequence learning contributes more, and states the transposed Transformer was validated as effective in Section V-D; the numerical ablation results are not included in the supplied excerpt.

Methodology in Plain English

The pipeline has four stages.

  1. Build the graph. Transaction records (timestamp, amount, sender, receiver) are crawled with the BlockchainSpider tool and turned into a multi-directed graph where nodes are Ethereum accounts and edges are transactions. Multiple edges can exist between the same pair of accounts, each carrying its own timestamp and amount.

  2. Sample with STRWalk. Instead of processing the whole network, the method walks outward from a starting account. Each step has two parts: a temporal step that weights candidate neighbors so that edges closer in time to the start node are more likely to be chosen (using an alias-sampling algorithm over a probability derived from timestamps), and a structural step that adds extra nodes from the same time interval to enrich the local structure. Time is bucketed into intervals computed in days (dividing by 86,400 seconds), which the authors say avoids both too-fine noise and too-coarse loss of signal. The walk repeats until a fixed length is reached or no neighbors remain, producing an ordered sequence of subgraphs.

  3. Encode each subgraph. For every subgraph in the sequence, node and edge features are assembled into matrices, neighbor features of differing sizes are aligned with a linear layer, LeakyReLU, and zero padding, and a graph attention network (GAT) computes attention scores over neighbors so the subgraph's central-node representation reflects which neighbors matter most. Running this across all intervals yields a temporal feature sequence.

  4. Learn the temporal evolution and classify. A transposed Transformer is used, which differs from a standard Transformer: rather than learning across time steps within an embedding dimension, it transposes the feature matrix so attention operates along the time axis for each embedding position, capturing how each feature changes over time. Masking prevents looking at future information; a feed-forward layer and a DNN (convolution and pooling layers plus a projection head) produce the final malicious/benign classification. Training minimizes cross-entropy loss plus a regularization term.

Why This Matters

Research impact: The paper argues that Web3 scams are a distinct phenomenon from phishing — differing in mimicry and track obscuration — and that prior work's focus on de-anonymization and phishing nodes leaves this gap open. It also contributes a temporal-subgraph-sequence formulation and an attribute analysis of a large labeled dataset, which the authors say was previously lacking for Web3.

Real-world applications:

  • Exchange and wallet risk screening, flagging accounts associated with drainer-style services before users interact with them.
  • Transaction monitoring for blockchain analytics firms and law enforcement tracing stolen funds through multi-transfer obfuscation.
  • Wallet-side warnings when a user is about to connect to or transact with a suspected scam service.
  • Auditing and threat intelligence for NFT airdrop and DeFi-style campaigns, the vectors the paper names as common concealment mechanisms.

Industry relevance: The scale problem is practical, not theoretical — with 2,762 million transactions, sampling-based pipelines like STRWalk are what make deployment feasible. The paper reports a concrete profit figure for one service, Venom Drainer, at over $27 million from about 15,000 victims, illustrating the economic incentive behind these scams. The authors include researchers from Tencent's Keen Security Lab, and the code and data are released publicly.

Future Directions

  • Cross-chain generalization: The framework is built and evaluated on Ethereum only; extending the temporal subgraph approach to other Web3 platforms is an open question.
  • Distinguishing scam subtypes: The paper treats Web3 scams as one class alongside phishing and normal accounts; separating Ponzi, rug pull, and drainer-style behaviors at the account level is a natural next step.
  • Sampling parameter sensitivity: STRWalk depends on the structure window size, temporal interval, and walk length, with the time unit fixed to a day; how sensitive detection is to these choices, and whether adaptive intervals help, is not resolved in the supplied text.
  • Adversarial adaptation: Since the paper's premise is that scammers actively obscure traces with mixers and multiple transfers, detection under deliberate evasion — and how quickly the model degrades as scam tactics shift — remains an open question.

Target Audience

Researchers and graduate students in blockchain security and graph machine learning; security engineers at exchanges, wallet providers, and blockchain analytics companies building account-risk systems; and anyone studying the boundary between phishing detection and broader Web3 fraud. Readers without background in GNNs, attention, or Transformers will find the methodology sections dense, though the problem framing and dataset analysis are accessible.

Authors’ abstract

The web3 applications have recently been growing, especially on the Ethereum platform, starting to become the target of scammers. The web3 scams, imitating the services provided by legitimate platforms, mimic regular activity to deceive users. However, previous studies have primarily concentrated on de-anonymization and phishing nodes, neglecting the distinctive features of web3 scams. Moreover, the current phishing account detection tools utilize graph learning or sampling…

The full abstract on arXiv
Read the original paper