Skip to content
Crypto.world

Research

The Anatomy of Address Poisoning on Ethereum: Funding Mechanisms, Scam Signatures, and Laundering via Tornado Cash

Address poisoning is a cheap phishing trick: a scammer creates an address whose first and last characters match an address the victim has already transacted with, then sends a tiny (often zero-value)…

The Anatomy of Address Poisoning on Ethereum: Funding Mechanisms, Scam Signatures, and Laundering via Tornado Cash
arXiv
2609.23405
Published
2026-09-20
Authors
Son Hoang Dau, Thanh Nguyen, Phuong Duy Huynh, Hong Yen Tran, Nicholas Huppert, Jeff Nijsse, Huong Ha, Xun Yi

AI summary

Overview

Research area: Blockchain security and crypto-asset forensics — specifically address-poisoning transfer (APT) phishing on Ethereum, its funding chain, its operational fingerprints, and the laundering of proceeds through Tornado Cash.

Technical level: Intermediate to Advanced. The paper assumes familiarity with Ethereum accounts, ERC-20 tokens, gas parameters, and transaction traces, and it uses graph-based clustering and information-theoretic consistency scores. The core ideas are still describable without that background.

Scope (one sentence): A large-scale on-chain study of roughly 56.4 million address-poisoning transfers on Ethereum between November 2022 and May 2025, covering how they are funded, what behavioural signatures scammers reuse, and how much scam money reaches Tornado Cash and other public services.

What This Paper Is About

Address poisoning is a cheap phishing trick: a scammer creates an address whose first and last characters match an address the victim has already transacted with, then sends a tiny (often zero-value) transfer from that lookalike address so it appears in the victim's history. When the victim later copies an address from that history, they may send real funds to the scammer. Prior work concentrated on detecting these transfers and counting victims; this paper instead asks who pays for the scam, what tell-tale habits scammers leave behind, and where the stolen money goes.

Key Contributions

  1. A finer taxonomy of the scam. The authors classify APTs into 11 subtypes along four axes — coin vs. genuine-token vs. fake-token, forward vs. backward, active vs. passive, and zero vs. nonzero value — replacing the three-type treatment in earlier literature, and they map which "master" roles (coin funder, token funder, coordinator) each subtype requires.

  2. Five families of scam signatures. They define signatures capturing (i) contracts created and invoked by scam addresses, (ii) APT-value patterns, (iii) relationships among funding values, APT values and transaction fees, (iv) gas configurations, and (v) use of public services, then test how consistently individual addresses reuse them and how well shared signatures corroborate known scam-operation links.

  3. The first study of Tornado Cash in the APT context. They examine how scammers use Tornado Cash both to fund APTs and to launder proceeds, propose APT-specific heuristics to group otherwise unrelated Tornado Cash users, and implement a tracing algorithm that follows individual tokens through swaps.

  4. First APT-proceeds estimate for public services. They estimate roughly $38.09M of APT scam proceeds entering Tornado Cash and $18.14M entering other public services over 11/2022–5/2025.

Main Findings

  • Scale of the dataset: 3,855,694 coin APTs, 14,799,949 genuine-token APTs and 37,708,070 fake-token APTs were identified; the victim dataset contains 4,209 victim transfers totalling around $166.31 million, of which 4,180 are type-1 and 29 are type-2.

  • Most APTs are passive and backward: 82.7% of coin APTs carry a nonzero value; among genuine-token APTs, 81.9% are zero-value, 80.3% are backward and 83.5% are passive; among fake-token APTs, 99.7% are backward and almost 100% are passive — meaning contracts, not the phishing addresses themselves, usually initiate them.

  • Fake tokens are heavily concentrated: 55,064 fake tokens were collected, the largest groups mimicking USDT (13,196, 24%), PEPE (7,452, 13.5%), USDC (6,106, 11.1%) and DAI (4,645, 8.4%).

  • Scam funding is industrialised: 57,558,073 master records across 5,307,343 unique transaction hashes were matched (2,780,856 coin-funding, 510,935 token-funding, 2,420,269 coordination). One coordinator, 0x49c6246ab5209733d2ed59af17f9081d6f138370, performed 12,639 coordinator transactions supporting 189,645 phishing addresses; another, 0xeFb3729Bc9C0ee64c718185d612C9538C0323306, appeared in 36,651 transactions supporting 162,479 phishing addresses. Of 51,414 unique masters, 43,906 appear in only one or two funding/coordinating transactions, and analysis focuses on the 7,508 involved in at least three.

  • Gas settings are a strong fingerprint: maxPriorityFeePerGas is the most consistent gas signature — 65.9% of masters use exactly one value, with mean consistency 0.81 (inverse-Simpson) and 0.8 (Shannon). Gas-signature cohesion is 0.58 for gasLimit, 0.88 for maxFeePerGas and 0.69 for maxPriorityFeePerGas.

  • Contract reuse is the strongest fingerprint of all: 4,112 masters called 6,176 private contracts; 76.5% called exactly one, about 90% scored at least 0.9 on both consistency measures (mean 0.95), and cohesion was 0.959 for the contract-call signature and 0.998 for the contract-generation signature.

  • A few large families dominate: 3,415 contract-creating masters form 38 small groups (size 2–31) plus four large groups of 1,699 (CG1), 467 (CG2), 254 (CG3) and 247 (CG4). These four supported 485,948 / 125,247 / 1,090,933 / 1,476,714 unique phishing addresses and 352,890 / 37,383 / 539,626 / 135,398 unique APT-carrying transactions; the remaining 534 masters each had a unique contract-generation signature.

  • Signatures agree across families: in CG1, all masters used exactly 3 for maxPriorityFeePerGas, with 1,643 of 1,699 also using 200 for maxFeePerGas and 18,500,000 for gasLimit; all 467 CG2 masters used the identical triple 3 / 200 / 18,500,000; 251 CG3 masters used 1.5 only, with three using both 0.1 and 1.5; CG4 was less uniform, with 149 masters using 1 and 73 using 1.5. Group cohesion was a perfect 1 for CG1, CG2 and CG3, and 0.91 for CG4.

  • Illustrative cluster (Figure 1): A distributor–master–phishing hierarchy where APT amounts sit in [10⁻⁶, 10⁻⁵); a phishing address 0x4f512487a746AB1638B5fFb0D8321dBDFA6Fb8eA received 50 ETH from a victim and reinvested it, while master 0x03CEc7583846EEEe04cDd0752A0B35bC8F793bff sent 10 ETH via Tornado Cash to distributor 0x306608Bf57e58A1263c491aD81b1A5fA2B420cfd, which then started a new cluster with an identical pattern.

  • Public services are gateways to the scam: among services used to fund masters, FixedFloat funded 291 masters, ChangeNOW 132 and WhiteBIT 39.

  • Victim losses skew to tokens: 327 ETH transfers (322 normal, 5 internal) and 1,505 genuine-token transfers exceeded $1,000 in historical value, corresponding to $12.21 million in ETH and $153.6 million in tokens; included are the 20 million USDT episode (frozen by Tether, 2023) and the 1,155 WBTC (~$70 million) case in 2024.

Methodology in Plain English

The authors pulled every Ethereum normal transaction, internal transaction and ERC-20 transfer event in their window from Etherscan and Google BigQuery — 1,399,654,309 normal and 639,829,207 internal transactions occupying 1.7 TB, plus 726 GB of token transfer data covering the top-100 tokens from CoinGecko and 55,064 identified fake tokens.

To spot an APT, they started from small genuine-token transfers (under $3) between two ordinary wallets. A transfer counts as an attack if one wallet's prior transaction history contains a real transfer to a different address that shares at least three leading and three trailing characters with the counterparty. Applying that rule across all three asset classes and recording whether the funds moved toward or away from the victim, who paid the fee, and whether the amount was zero, produces the 11-subtype labelling.

With attacks labelled, they worked backwards to the money. For each phishing address they walked its APTs oldest to newest and matched each one to the nearest incoming non-APT coin or token transfers that together covered the APT amount plus gas — allowing partial funding and re-use of one funding transfer across several APTs. Wallet addresses that pay fees themselves were recorded as coordinators. This produced the master/distributor hierarchy.

They then measured how stereotyped scammer behaviour is. A consistency score (inverse-Simpson and Shannon, over the values an address uses) asks whether one address keeps repeating the same gas settings or contracts. A cohesion score asks the reverse question: given a group of masters sharing a signature value, what fraction of pairs are already connected in a "guilt-by-association" graph where edges mean shared victims, mutual funding, or a common non-service funder. High cohesion means the signature recovers real operational links rather than coincidence. Contracts are compared by hashing their bytecode and truncating to 40 hex characters, avoiding pairwise code comparison.

For Tornado Cash, they decoded deposit and withdrawal calls across TC pools and routers (36,435 deposits, 35,076 withdrawals, with ETH pools accounting for about 98% of activity) and used the master/distributor hierarchy as a bridge between otherwise unlinkable depositors and withdrawers.

Why This Matters

The paper moves APT research from "detect the transfer" to "map the organisation," which matters because the transfers themselves are trivial and cheap while the infrastructure behind them is where scale and attribution live. Showing that a handful of contract-derived groups supported over a million unique phishing addresses each reframes APT as an industrial operation with toolkits, not thousands of independent opportunists.

Real-world applications:

  • Exchange and wallet risk screening: gas configurations and bytecode-hash contract signatures can flag addresses that detection-by-transfer alone would miss, including freshly created phishing addresses that inherit a known family's fingerprint.
  • Law-enforcement tracing: the master/distributor hierarchy plus the token-following tracer gives investigators entry points into mixer traffic that address-level graph analysis does not provide.
  • Victim-side wallet UX: the finding that nearly all fake-token APTs are passive and that most genuine-token APTs are zero-value suggests filtering should key on contract-initiated log entries, not just visible transfers.
  • Service-provider compliance: identifying which services (FixedFloat, ChangeNOW, WhiteBIT among them) fund masters, and that $38.09M in proceeds entered Tornado Cash, gives concrete targets for monitoring.

Industry relevance: this directly supports blockchain-analytics vendors, custodial exchanges, and compliance teams, and it formalises the "attack strategies" that earlier work (Tsuchiya et al., USENIX'25) described only qualitatively.

Future Directions

  • Extend the signature catalogue to other chains and to non-Ethereum mixers, since the current window is 11/2022–5/2025 and Ethereum-only, and ETH pools account for about 98% of the Tornado Cash activity observed.
  • Test whether the four large contract groups (CG1–CG4) are separately operated or one actor, given that their gas profiles differ (3/200/18,500,000 versus 1.5 versus a mix of 1 and 1.5) even though each group is internally near-perfectly cohesive.
  • Resolve the sub-3-USD and three-character matching thresholds: the detection rules are heuristics, and the paper does not report a precision/recall evaluation against a labelled ground truth, so the sensitivity of the 56.4 million APT count to those thresholds is an open question.
  • Close the gap between proceeds entering services and proceeds recovered: the paper estimates flows into Tornado Cash ($38.09M) and other public services ($18.14M) but reports only one freeze (20 million USDT by Tether), leaving the question of what fraction of $166.31M in victim transfers is realistically recoverable.

Target Audience

Blockchain-forensics and security researchers building on the APT detection line (Ye et al. WWW'24, Guan-Li CCS'24, Chen et al. NDSS'25, Tsuchiya et al. USENIX'25); anti-money-laundering and compliance analysts at exchanges and analytics firms; and graduate students in crypto-asset security who want a worked example of combining large-scale on-chain data engineering with graph clustering and behavioural fingerprinting.

Note: the supplied text is truncated partway through Section 5.5, so the full definitions of the funding-residual signature family and the detailed results of the Tornado Cash grouping and tracing in Section 6 are not available in the content above; only the summary-level claims from the abstract and contributions list are reported here.

Authors’ abstract

Address-Poisoning Transfer (APT) is a prevalent blockchain phishing scam in which a scammer poisons a victim's address book by generating a transfer with a phishing address that looks similar to a benign address that the victim has previously interacted with. Although simple, APT phishing attacks have cost users millions of dollars in recent years, which has captured the attention of the research community (Ye et al. WWW'24, Guan-Li CCS'24, Chen et al. NDSS'25, Tsuchiya et…

The full abstract on arXiv
Read the original paper