Skip to content
Crypto.world

Research

Phishing Detection in Ethereum via Temporal Graph Contrastive Learning

Ethereum phishing — where attackers impersonate legitimate entities to trick users into sending crypto, then rapidly launder it — caused over $1B in losses across 296 incidents in 2024 alone…

Phishing Detection in Ethereum via Temporal Graph Contrastive Learning
arXiv
2605.01207
Published
2026-05-02
Authors
Cong Wu, Jing Chen, Siqi Lin, Hongda Li, Ziming Zhao

AI summary

Overview

Research area: Blockchain security / graph machine learning — specifically phishing detection on the Ethereum network using temporal graph neural networks and self-supervised contrastive learning.

Technical level: Advanced. The paper assumes familiarity with graph neural networks, attention mechanisms, memory networks (GRU), and contrastive/self-supervised learning objectives.

Scope (one sentence): The paper presents PhishEye, a self-supervised system that models Ethereum transactions as a heterogeneous temporal attributed multi-graph and uses a temporal graph contrastive learning model (PhishTGL) to flag phishing accounts and transactions, evaluated on a 161,658-address dataset and in a 15-month live deployment.

What This Paper Is About

Ethereum phishing — where attackers impersonate legitimate entities to trick users into sending crypto, then rapidly launder it — caused over $1B in losses across 296 incidents in 2024 alone, according to the paper. Existing graph-based detectors treat the transaction graph as static (missing how phishing tactics evolve) or "semi-dynamic" (bolting an LSTM onto a static GNN, which assumes evenly spaced events and misses the bursty, irregular timing of blockchain activity); they also flatten diverse transaction types (Ether, fungible tokens, NFTs) into one homogeneous graph and depend on scarce labeled data. The goal is a fully dynamic, self-supervised detector that captures both the timing and the heterogeneity of on-chain activity.

Key Contributions

  1. PhishEye system: The authors describe it as the first self-supervised system for phishing detection on Ethereum, built on heterogeneous temporal graph contrastive learning and able to model Ether, FT, and NFT transfers as distinct transaction types.
  2. PhishTGL model: A novel temporal graph learning model combining learnable functional time encoding, a GRU-based node memory for historical interactions, temporal multi-head attention over neighbors, edge representation learning, and node-level contrastive learning between perturbed graph views.
  3. Comprehensive evaluation plus real-world deployment: Experiments on a compiled dataset and a live deployment on Ethereum (May 1, 2023 – July 31, 2024), including analysis of fund flows of newly detected phishing addresses to characterize post-attack behavior.
  4. Public dataset: 1,000 phishing and 160,658 non-phishing addresses, with 48,135 phishing and 368,406 non-phishing transactions, plus open-sourced code at an anonymous.4open.science repository (PhishEye-3CEE).

Main Findings

  • Transaction-level detection: PhishEye achieved an F1 score of 87.23% and an AUC of 98.43% for phishing transaction detection on the compiled dataset.
  • Account-level detection: It achieved an F1 score of 94.19% and an AUC of 98.03% for phishing account detection.
  • Live deployment volume: Between May 1, 2023 and July 31, 2024, the deployed prototype detected 2,153 phishing addresses, of which 1,803 were identified before any public reports and later confirmed via victim disclosures and phishing site analysis. Another 196 addresses were flagged as suspicious and validated by risk auditors.
  • Loss prevention: The system helped alert to losses of $2,039.1M out of an estimated $2,813.7M in phishing-related losses during that period; the abstract states alerts helped prevent losses exceeding 2 billion USD.
  • Error rates versus prior methods: In deployment PhishEye reported a false negative rate of 1.31% and a false positive rate of 2.75%, compared with TTAGN (19.65%, 26.19%), TGC (18.72%, 32.14%), and Trans2vec (11.37%, 19.43%).
  • Phishing address structure: The average in-degree, out-degree, and total degree of phishing addresses in the dataset were 22.63±29.16, 22.63±24.43, and 51.75±29.42 respectively (Figure 3).
  • Motivating pattern: A March 29, 2023 incident shows a phisher address (0xe9..9c) receiving a burst of transfers from multiple victims, mixing fungible and non-fungible tokens, then rapidly funneling assets to another address (0xa3..92) to obscure the trail.
  • Comparison baselines: DeepWalk, GraphSAGE, and TGAT are named as comparison graph models, but their quantitative results are cut off in the supplied text (the paper content truncates mid-sentence in Section 4.1.1), so specific per-baseline scores are not reported here.

Methodology in Plain English

The system watches Ethereum transactions and turns them into a graph where accounts are nodes and transfers are edges. Unlike prior work, it keeps three things that matter: (1) time — every edge carries a timestamp, and the graph keeps growing as new blocks arrive; (2) type — Ether transfers, fungible-token transfers, NFT transfers, contract interactions, and internal contract transactions are treated as different kinds of edges rather than all the same; and (3) rich account descriptions — because Ethereum accounts carry little inherent information, the authors hand-build 18 features (nine about transaction activity such as in/out degree and value flows, nine about network position such as PageRank and clustering), plus identifiers for whether an account is user-controlled (EOA) or a contract (CA).

The learning model, PhishTGL, works in layers. A mathematical time encoding turns timestamps into vectors the network can learn from. Each account keeps a memory of its past interactions, updated with a GRU every time it transacts. When producing an account's representation, a multi-head attention mechanism weighs its neighbors based on their relevance and timing. Edges get their own representations too — notably, the model aggregates prior incoming transactions to the same recipient, since a burst of incoming transfers to one address is a phishing signature.

Training needs no labels: the graph is perturbed twice (20% of edges randomly removed, 20% of node features randomly masked) to create two views, and the model is trained to make the same node's representations agree across views while pushing other nodes apart (a contrastive loss with temperature τ = 1). Only afterward is a lightweight classifier (LightGBM) trained on the learned representations to output phishing/non-phishing labels for nodes and edges. Hyperparameters include 2 graph layers, 8 attention heads, 128-dimensional memory and temporal representations, learning rate 0.0001, batch size 256, a 7:3 train-test split, and 10-fold cross-validation, with random edge upsampling and node downsampling at a 0.8 ratio to handle class imbalance.

Why This Matters

Impact on research: The paper argues that static and semi-dynamic graph learning are structurally mismatched to blockchains, where events are irregular and bursty rather than uniformly spaced. It offers a template — heterogeneous temporal multi-graph plus self-supervision — for label-scarce, continuously evolving financial graphs, and contributes a labeled benchmark (1,000 phishing / 160,658 non-phishing addresses) that others can reuse.

Real-world applications:

  • Wallet and exchange risk engines that warn users before they sign a transfer to a flagged address.
  • Block explorers and threat-intelligence feeds that label addresses and transactions in near real time.
  • Law-enforcement and incident-response tracing, aided by the paper's fund-flow analysis of newly detected phishing addresses.
  • Custody and compliance platforms screening counterparties before large transfers.

Industry relevance: The deployment numbers — 1,803 addresses flagged before public disclosure, and alerts tied to $2,039.1M of an estimated $2,813.7M in losses — speak directly to security vendors, exchanges, and DeFi protocols that need early warning rather than post-hoc labeling. The work was conducted with two anonymized companies and accepted at ACM CCS 2026.

Future Directions

  • Extending beyond Ethereum: the HTAMG formulation is chain-agnostic in principle, but the paper evaluates only Ethereum; cross-chain generalization (and cross-chain laundering, e.g., via bridges) is untested.
  • Adversarial robustness: the contrastive views come from random edge removal and feature masking; how the detector holds up against attackers who deliberately reshape transaction patterns to evade the model is not addressed in the reported results.
  • Label efficiency at scale: the paper uses only 1,000 verified phishing addresses; whether performance holds as positive labels grow noisier, or whether the self-supervised representations transfer to new phishing families with zero labels, remains open.
  • Latency and streaming behavior: the deployment reports detection volume and error rates over 15 months, but the supplied content does not report per-transaction latency or how quickly the graph and node memory update under mainnet throughput — a natural next question for real-time alerting.

Target Audience

Researchers and graduate students in blockchain security, graph machine learning, and financial fraud detection; security engineers at exchanges, wallets, and threat-intelligence vendors who need practical on-chain detection; and practitioners interested in self-supervised learning applied to dynamic, label-scarce, heterogeneous graphs. Readers without a background in GNNs and contrastive learning will find the modeling sections demanding.

Authors’ abstract

Blockchain and decentralized finance have revolutionized the financial ecosystem while simultaneously exposing it to cryptocurrency phishing attacks. Existing phishing detection methods primarily rely on graph learning, but they face significant limitations. Static graph learning approaches fail to account for the temporal evolution of phishing patterns, while semi-dynamic methods, such as those combining static GNNs with LSTM, struggle to capture the irregular and bursty…

The full abstract on arXiv
Read the original paper