Transformer-based comparative multi-view illegal transaction detection

Meiling Zhong, Yong Wang, Jingke Yan, Yao Cheng, Peihao SunPLoS 2023 · DOI · OpenAlex

Note on Publication Status: This article was originally published in PLOS ONE in January 2023 but has since been retracted. The summary below reflects the technical claims and methodology presented by the authors in the original manuscript, though the validity of the results should be viewed with caution given the retraction.

Abstract

In recent years, as the Ethereum platform has grown, numerous individuals have used illegal transactions to defraud investors, causing billions of dollars in losses. Current detection methods often rely on a single view—either the smart contract's code or account transaction features—which fails to capture the global structure and semantic features between tokens. This paper investigates a Transformer-based model for contrasting illegal transaction detection networks under multiple views (TranMulti-View Net). The model utilizes Transformers to learn multi-view fusion representations, maximizing the interaction information between different features. By capturing remote dependencies in token sequences and sharing features across views, the model aims to identify important semantic information. Furthermore, the authors employ semi-supervised contrastive learning, which reportedly outperforms direct fusion schemes. Experimental results claimed a precision score of 98% in detecting illegal transactions.

Introduction

As the Ethereum ecosystem expands, financial crimes such as phishing, money laundering, and smart contract fraud have become increasingly sophisticated. Traditional detection frameworks often suffer from "tunnel vision," focusing exclusively on either contract code (logic) or transaction history (fund flow).

This paper argues that single-view approaches are inherently incomplete. A malicious contract might possess benign-looking code while exhibiting suspicious transaction patterns, or vice versa. Moreover, standard sequential models often fail to capture "remote dependencies"—the long-distance relationships between tokens in code or steps in a transaction history that signal fraudulent intent. The proposed TranMulti-View Net seeks to bridge this gap by integrating these disparate data sources into a unified representation.

Methodology: TranMulti-View Net

The core of the framework is a multi-view fusion architecture that leverages several modern deep learning techniques:

  1. Transformer-Based Feature Extraction: Moving beyond traditional RNNs or CNNs, the model employs Transformers to process sequences derived from both contract code and transaction data. This architecture is specifically chosen for its ability to capture global structural information and long-range semantic dependencies within the data.
  2. Multi-View Fusion: The model processes two distinct inputs simultaneously:
  3. Contract Code View: Tokenized representations of the smart contract's source code or bytecode.
  4. Account Transaction View: Sequential data representing the history of interactions and fund movements.
  5. Contrastive Learning: Rather than simple concatenation, the authors use a semi-supervised contrastive learning approach. This forces the model to learn a shared representation space where the code view and transaction view of the same entity are closely correlated, helping the model identify the underlying semantic "signatures" of illegal activity.

Key Insights and Results

The authors emphasize the power of "mutual sharing" between views. By allowing the model to share information across the code and transaction domains, it can effectively "fill in the gaps" when one view is obfuscated or sparse. For instance, if a contract's code is intentionally complex to hide its purpose, its transaction patterns may still reveal fraudulent behavior.

The study concludes that semi-supervised contrastive learning significantly outperforms direct fusion (simply joining two feature vectors). This method ensures the model understands the functional relationship between code logic and actual behavior. Based on these techniques, the authors reported a 98% precision rate, claiming that the multi-view approach drastically reduces false positives compared to single-view benchmarks.

Limitations and Retraction

Despite the high performance reported, several critical factors must be considered:

  • Retraction Status: The official retraction of the paper implies significant concerns regarding data integrity or methodology discovered post-publication.
  • Computational Complexity: Transformers require substantial computational resources, and the paper does not fully address the latency requirements for real-time monitoring on the blockchain.
  • Adversarial Evolution: As detection models evolve, attackers may develop "cross-view" obfuscation techniques specifically designed to deceive fusion-based models.
  • Data Verification: Due to the retraction, the original data availability statement was removed, making independent replication and verification of these results difficult for the research community.

Summary updated

Abstract

In recent years, as the Ether platform has grown by leaps and bounds. Numerous unscrupulous individuals have used illegal transaction to defraud large sums of money, causing billions of dollars of losses to investors worldwide. Facing the endless stream of the illegal transaction based on Ether smart contracts problems, such as illegal transaction, money laundering, financial fraud, phishing. Currently, illegal transaction are only detected by a single view of the smart contract's contract code view feature and account transaction view feature, which is not only incomplete, but also not fully representative of the smart contract's features. More importantly, the single view detection model cannot accurately capture the global structure and semantic features between the Tokens of the view features. In this case, it is particularly important that all view features are shared among themselves. In this paper, we investigate a Transformer-based model for contrasting illegal transaction detection networks under multiple views (TranMulti-View Net). The model in this paper is based on Transformer to learn a multi-view fusion representation, which aims to maximise the fusion of the interaction information of different view features under the same condition. In this model we first use the Transformer model to learn global structure and semantic features from a sequence of Tokens tokenised by a view, capturing the remote dependencies of Tokens in the view features, and then we share the contract code view features and the account transaction view features across all views to learn important semantic information between views from each other. In addition, we find that the approach of semi-supervised training of multi-view features using contrast learning outperforms the scheme of prediction based on direct fusion of different view features, resulting in stronger correlation between view features. As a result, the underlying semantic information can be captured more accurately, leading to more accurate predictions of illegal transaction. The experimental results show that our proposed TranMulti-View Net obtains good detection results with a Precision score of 98%.
BibTeX
@inproceedings{zhong2023transformer,
  author = {Meiling Zhong and Yong Wang and Jingke Yan and Yao Cheng and Peihao Sun},
  title = {{Transformer-based comparative multi-view illegal transaction detection}},
  booktitle = {PLoS},
  year = {2023},
  doi = {10.1371/journal.pone.0276495},
}
Figure