# Beyond the Hype: A Large-Scale Empirical Analysis of On-Chain Transactions in NFT Scams

Crypto.world

> NFTs have created a large digital-asset market, and with it a wave of phishing scams that steal wallets and assets — yet prior work has focused on vulnerabilities in smart contract code, rug pulls…

**arXiv**: 2512.01577

**Published**: 2025-12-01

**Authors**: Wenkai Li, Zongwei Li, Xiaoqi Li, Chunyi Zhang, Xiaoyan Zhang, Yuqing Zhang

## AI summary

## Overview

**Research area:** Blockchain security, specifically NFT phishing fraud analysis using on-chain transaction graph analysis.

**Technical level:** Intermediate — assumes familiarity with Ethereum, ERC20/ERC721/ERC1155 token standards, and basic graph concepts (nodes, edges, connected components), but the analysis is descriptive rather than algorithmically heavy.

**Scope (one sentence):** The paper builds a transaction graph from Ethereum data covering the first 18 million blocks and empirically characterizes how NFT phishing accounts are distributed, which token standards they use, and how they interact with normal accounts.

## What This Paper Is About

NFTs have created a large digital-asset market, and with it a wave of phishing scams that steal wallets and assets — yet prior work has focused on vulnerabilities in smart contract code, rug pulls, and wash trading rather than on how NFT phishing actually looks in the on-chain transaction graph. The authors address this gap by collecting NFT phishing reports and address labels from multiple platforms, standardizing the transaction records across ERC20, ERC721 and ERC1155, and constructing a graph in which accounts are nodes and transactions are edges. The goal is to describe the distribution, transaction features, and interaction patterns of NFT phishing so that future detection methods have a behavioral reference.

## Key Contributions

1. **First empirical analysis of the NFT phishing scam graph.** The authors state this is the first study to systematically examine NFT phishing through graph analysis, and argue it demonstrates that NFT phishing can be meaningfully studied with graph-theoretic methods.
2. **A quantitative distribution finding.** They show that while normal transactions account for 96.71% of all transactions, phishing addresses are disproportionately present relative to their share of accounts.
3. **Evidence of coordinated, targeted behavior.** They report that NFT phishers act in coordinated groups aimed at specific account targets rather than collaborating heavily among themselves.
4. **Evidence of multi-protocol evasion and shorter cycles.** They find that fraudsters tend to run mixed multi-protocol transactions with the same victim and that phishing transaction cycles are shorter than standard transactions, with more multi-party transactions involved.

## Main Findings

- **Phishing addresses are few but appear in many scenarios.** In the collected data there are 37 NFT phishing accounts versus 3,898 normal accounts, so phishing accounts are only 0.94% of total accounts; nevertheless they appear in 8.36% of transaction scenarios.
- **Interaction rises sharply at scale.** The proportion of mixed graphs (containing both phishing and normal addresses) goes from 2.74% of small graphs to 10.55% of large graphs, and the paper states the interaction probability with normal addresses increases 2.85 times in large-scale transaction graphs.
- **Most graph structure is legitimate.** Of the 3,925 weakly connected components found, 275 met the minimum-10-node threshold, containing 3,935 nodes in total: 246 normal graphs and 29 mixed graphs. 89.45% of the large graphs transfer only to normal addresses.
- **Phishing nodes cluster in mixed graphs.** Although only 37 phishing nodes exist in the graphs, 78.37% of them appear inside mixed graphs, indicating contact between phishing units and ordinary users.
- **ERC20 dominates edge counts.** In normal graphs ERC20 accounts for 54.30% of edges and in mixed graphs 76.34%, far above ERC721 (35.33% normal, 33.18% mixed) and ERC1155 (17.20% normal, 7.53% mixed). Only the ERC20 share exceeds 50%, which the authors read as NFT transactions usually being accompanied by ERC20 transactions.
- **ERC721 transactions carry more NFTs per edge.** The average NFT amount for ERC721 is 1.19 in normal graphs and 1.48 in mixed graphs, slightly higher in mixed transactions.
- **Cross-protocol activity is higher for phishers.** At least two protocols are used in 2.27% of normal edges versus 3.38% of mixed edges. Broken down: ERC20+ERC721 is 58 (1.80%) normal versus 18 (2.76%) mixed; ERC20+ERC1155 is 10 (0.31%) versus 3 (0.46%); ERC721+ERC1155 is 15 (0.47%) versus 1 (0.15%); all three protocols appear in 5 (0.16%) normal edges and 0 (0.00%) mixed edges.
- **Phishers target normal accounts, not each other.** In the mixed graph's 651 transactions, normal-to-normal is 435 (66.82%), phishing-to-normal is 117 (17.97%), normal-to-phishing is 95 (14.59%), and phishing-to-phishing is only 4 (0.61%).
- **Only a small share of nodes drive token flow.** Convergence nodes (high in-degree, low out-degree) are 62 (1.57%), distribution nodes (low in-degree, high out-degree) are 80 (2.03%), and bidirectional hubs are 553 (14.05%).

## Methodology in Plain English

The authors first gather data in the style of prior NFT phishing work: incident reports and security write-ups from Chainabuse, Twitter/X, SlowMist, PeckShield and PANews, plus address labels from ScamSniffer, Etherscan and Chainabuse. They then pull the corresponding on-chain transaction and log data for the first 18 million Ethereum blocks using a tool called BlockchainSpider.

Because different token standards store transaction fields differently, all three token types (ERC20, ERC721, ERC1155) go through the same four-stage pipeline described in their Algorithm 1: load the normal-account data, load the phishing-account data, align the fields to a single unified field set (filling missing fields with `None` and reordering columns), then merge the two datasets and convert types (block numbers to integers, NFT timestamps to datetimes, amounts to numeric). Each record is tagged as "normal" or "phishing" and annotated with its protocol type.

Next they build the graph. Accounts are nodes and transactions are edges, modeled as an undirected graph so that the focus is on association rather than fund direction, using the cuDF and cuGraph libraries. To keep the analysis meaningful they drop system-level addresses (zero address, dead address) and filter out "super hub" nodes whose degree exceeds a threshold set to 30 by default. Multiple transactions between the same pair of addresses are collapsed into one edge per protocol type, and each edge gets features such as `tx_count`, `total_amount`, `avg_amount`, `first_block`, `last_block`, `block_span`, `token_contract_count` and `series_count`. Edge tables from the three protocols are joined so an edge carries both protocol-specific and cross-protocol statistics.

Finally, a weakly connected components algorithm extracts groups of interrelated addresses, which are labeled phishing graph, normal graph, or mixed graph depending on the node labels inside. Components with fewer than 10 nodes (default) are discarded as isolated or too small to be informative.

## Why This Matters

**Impact on research.** Prior NFT scam research concentrated on rug pulls, wash trading, and promotion fraud, and prior phishing work built classifiers on transaction and NFT features. This paper supplies the missing descriptive layer — what phishing actually looks like as a graph — which is the kind of ground truth needed to design graph-based detectors and to argue which structural features are worth engineering.

**Real-world applications:**

- **Exchange and wallet risk screening:** flag accounts whose graph position resembles the phishing patterns reported here, such as short-cycle multi-protocol interactions with a single counterparty.
- **Victim tracing and fund-flow investigation:** the finding that phishing nodes concentrate in mixed components, and that mixed components grow more common at larger scale, gives investigators a prioritization signal for which clusters to examine.
- **Protocol and marketplace monitoring:** the heavy co-occurrence of ERC20 transfers alongside ERC721/ERC1155 transfers in mixed graphs suggests monitoring rules that watch approval and token-transfer combinations rather than NFT transfers alone.
- **Airdrop and promotion campaign vetting:** the motivating case in the paper is a phishing campaign hidden behind promotional NFT airdrops that abused the Seaport protocol to bypass detection, which is a direct warning for teams running such campaigns.

**Industry relevance.** The paper cites a market projected to exceed 26.41 billion USD by 2024, Elliptic's finding of over $100 million in illicit NFT scam gains between July 2021 and July 2022 at roughly $300,000 per incident, and a 2023 campaign in which minors were enlisted for large-scale phishing that compromised over 32,000 wallets and 900 Discord servers and stole assets valued at $73 million. Those figures frame NFT phishing as a financially material operational risk, not a niche curiosity.

## Future Directions

- **Turn the descriptive features into a detector.** The paper explicitly positions its behavioral features as a reference for future graph-based NFT phishing detection; no detection model or accuracy is reported here, so building and benchmarking one is the obvious next step.
- **Explain the multi-protocol behavior mechanistically.** The data show phishers mix ERC20 with ERC721/ERC1155 more often than normal users, and the authors interpret this as spreading the risk of being traced, but the underlying approval and transfer mechanics are not traced out transaction by transaction.
- **Expand beyond the current scope.** The dataset covers 37 phishing accounts and the first 18 million Ethereum blocks; extending to more phishing labels, other chains in the multi-chain NFT ecosystem (the paper names Solana, Cardano, EOS, Tezos and Polygon), and later blocks would test whether the patterns hold.
- **Reconcile the stated research questions.** The paper says the investigation is structured around four central research questions but enumerates only RQ1 through RQ3, and its analysis is descriptive without reported statistical significance testing — a natural follow-up is confirming whether the observed differences (for example the 2.27% versus 3.38% cross-protocol rates) are statistically robust.

## Target Audience

Blockchain security researchers and graduate students studying fraud detection and transaction-graph analytics; security engineers at exchanges, wallet providers, and NFT marketplaces who build risk-screening rules; and incident responders and on-chain investigators who trace stolen NFT and token flows. Readers wanting a detection algorithm or measured classifier performance will not find it here — the value is in the empirical characterization of NFT phishing behavior, so readers without a security background can follow the findings but will need some grounding in Ethereum token standards to interpret the protocol-level results.

## Authors’ abstract

Non-fungible tokens (NFTs) serve as a representative form of digital asset ownership and have attracted numerous investors, creators, and tech enthusiasts in recent years. However, related fraud activities, especially phishing scams, have caused significant property losses. There are many graph analysis methods to detect malicious scam incidents, but no research on the transaction patterns of the NFT scams. Therefore, to fill this gap, we are the first to systematically…

[The full abstract on arXiv](https://arxiv.org/abs/2512.01577)

[Read the original paper](https://arxiv.org/abs/2512.01577)

---

Canonical: https://crypto.world/paper/beyond-the-hype-a-large-scale-empirical-analysis-of-on-chain-2512.01577
