Skip to content
Crypto.world

Research

MixGuard: Towards Detecting and Understanding Mixer Laundering on Ethereum

Mixers such as Tornado Cash and Railgun hide the link between deposits and withdrawals, which protects legitimate privacy but also lets attackers move stolen funds.

MixGuard: Towards Detecting and Understanding Mixer Laundering on Ethereum
arXiv
2609.27807
Published
2026-09-15
Authors
Qishuang Fu, Hang Zheng, Xihan Xiong, Joseph K. Liu, Yixin Liu, Shirui Pan, Qin Wang, Weiqing Wang, Zhipeng Wang, Tsz Hon Yuen

AI summary

Overview

Research area: Blockchain security and anti-money laundering (AML) on Ethereum, specifically the use of privacy mixers (Tornado Cash and Railgun) to launder illicit funds.

Technical level: Intermediate. Readers need basic familiarity with Ethereum transactions, smart contracts, and graph-based machine learning, but the measurement sections are accessible to a general blockchain audience.

Scope: The paper builds the first public case-level dataset of real-world mixer laundering, measures what distinguishes laundering from ordinary mixer usage, and proposes a detection-and-grouping framework evaluated on unseen cases and live Ethereum data.

What This Paper Is About

Mixers such as Tornado Cash and Railgun hide the link between deposits and withdrawals, which protects legitimate privacy but also lets attackers move stolen funds. Existing work either studies mixers for deanonymization or studies general Ethereum money laundering, but neither provides labeled, case-level evidence of laundering that passes through a mixer. The authors fill that gap by assembling real attack cases into a labeled dataset, characterizing how laundering differs from background mixer activity, and turning those observations into a detector that flags laundering-related mixer transactions and groups them by case.

Key Contributions

  1. First public case-level mixer-laundering dataset. MixLaunder covers 27 real-world attack cases involving Tornado Cash and Railgun from 2020 to 2025, labeling 9,300 laundering mixer transactions (5,349 deposits and 3,951 withdrawals) with case identities and traced upstream and downstream fund flows, including deposits totaling approximately $1.1 billion.
  2. Systematic measurement grounded in real cases. The authors compare laundering transactions against same-day background mixer usage and identify five recurring laundering strategies spanning fund preparation, mixer use, post-withdrawal restructuring, and intra-case coordination.
  3. Evaluation of existing countermeasures. The paper documents coverage gaps in Tornado Cash's Compliance Tool, Railgun's Private Proof of Innocence screening, and three representative address-level deanonymization heuristics, the broadest of which covers only 4 of the 27 cases.
  4. A measurement-guided detection and grouping framework. MixGuard combines a tri-view encoder (address, fund-flow path, local topology) with two-stage grouping to perform transaction-level laundering detection and case-aware grouping, achieving 97.89% detection precision, 98.73% group purity, and 95.09% average top-ten case coverage under case-level holdout, and surfacing four validated incidents involving more than $15 million in January 2026 Ethereum data.

Main Findings

  • Role separation at the mixer boundary: In 24 of the 27 cases no address serves both deposit and withdrawal roles, and only 1.48% of laundering addresses serve both. Laundering deposit addresses act as aggregation inlets with high in-degree and low out-degree, while 98.6% of laundering withdrawal addresses have at most one day of on-chain activity before their first mixer interaction.
  • Concentrated behavior: Tornado Cash interactions and direct transfers make up 97.8% of dominant transaction types for laundering addresses, whereas background mixer addresses spread activity across Tornado Cash, direct transfer, loop, Railgun, split, and other types.
  • Rapid, interleaved mixer use: About 85.5% of intervals between temporally adjacent deposits and withdrawals are within one day and 95.8% within seven days, with the tail beyond one week below 5%; the median interval falls within one day in 19 of 27 cases. 25 cases show interleaved deposit and withdrawal phases, with only 2 strictly sequential.
  • Slow, concentrated upstream preparation: Only 5.2% of laundering deposits are supplied by multiple sources at the immediate upstream hop, versus 30.3% for background deposits, and in 20 of 27 cases no more than 10% of deposits have multiple immediate upstream sources. The median time from source to mixer is 10.5 days for illicit funds versus 2.8 hours for background funds, and upstream behavior is dominated by direct transfer (88.8%), with swap at 8.7%.
  • Post-withdrawal restructuring: The median time from withdrawal to first service destination is 6.7 days for illicit funds versus 2 hours for background. Downstream, swap accounts for 39.9% (versus 1.6% in background), split 26.2%, aggregation 12.7%, and direct transfer only 16.7%; these three restructuring behaviors account for a median of 64.2% of downstream activity and exceed direct transfers in 19 of 27 cases.
  • Locally tight, globally loose coordination: On average, 90.1% of deposit addresses in a case have a deposit whose immediate upstream sender also funds another deposit address, and 68.1% of withdrawal addresses show the analogous downstream overlap; the most active 1-hour window per case contains on average 33.3% of its deposit transactions. Yet deposit activity spans a median of 18.8 days (longest 329 days) across an average of 35.7 distinct 1-hour windows, and among cases with multiple windows the median share of window pairs sharing no mixer-side address is 68.0% for deposits and 67.6% for withdrawals.
  • Weak protocol-level screening: Tornado Cash's optional Compliance Tool only lets depositors prove deposit–withdrawal links and performs no risk assessment; in the Harmony case, 4 withdrawal addresses used only 2 relayers across 8 withdrawals with an average relayer fee of about 81 ETH per withdrawal, leaving 19 ETH to the recipient, without any risk alert.
  • Gaps in Railgun's deposit-only screening: Across the three Railgun cases, 28 of 37 laundering deposit addresses (75.68%) had no deposits flagged; of the remaining nine, seven had all deposits flagged, while two received mixed verdicts. 18.92% of deposit addresses had all deposits flagged and 5.41% had earlier deposits unflagged before later ones were banned; two addresses (0x6c4f and 0xea5f) received further withdrawals before their flagged deposits were refunded.
  • Deanonymization heuristics barely fire: Double role hits 1.48% of addresses and 3 of 27 cases, direct linkage 1.27% and 4 of 27 cases, and deposit-as-initiator 0.64% and 2 of 27 cases.
  • Detection performance: Under strict case-level holdout, MixGuard reaches 97.89% detection precision and 98.73% group purity, with its top ten groups covering 95.09% of each case's transactions on average, outperforming representative baselines.

Methodology in Plain English

The authors work measurement-first. They start by collecting publicly documented attack cases from incident repositories and on-chain investigator write-ups, then verify each case against real Ethereum data by matching the reported clues to actual Tornado Cash or Railgun deposit and withdrawal transactions. A case is kept only if it is a publicly disclosed real attack, is reported to have used one of the two mixers, and yields verifiable deposit and withdrawal records.

Because a single raw transaction can bundle router calls, relayer payments, helper contracts, and token transfers, they extract one standardized record per mixer interaction capturing the transfer that actually moves funds into or out of the mixer. From each such record they trace funds outward: backward from deposits and forward from withdrawals, keeping only transactions that are temporally consistent and can carry the traced asset and amount, and following funds through swaps via event logs. Tracing stops when nothing attributable remains, no related transaction exists, or funds reach a labeled service such as an exchange.

To create a fair comparison, they build a background set from the same days each case used the mixer, with all labeled laundering transactions removed. They then compare laundering and background activity across address behavior, deposit–withdrawal timing, fund flows, and coordination, which produces the five strategies. Guided by those strategies, MixGuard represents each mixer transaction through three views – an address view and a path view encoded by MLPs, and a topology view using a two-layer GraphSAGE over the local fund-flow subgraph, with mean-pooled subgraph features concatenated to the node embedding so the transaction's own signal is not diluted – and fuses them. Two-stage grouping first forms small, reliable "core" clusters from strong local ties, then merges related cores using learned weaker associations, which counters the fragmentation that pure address-level linking produces. The provided excerpt ends during the encoder fusion description, so the remaining architectural and experimental details (fusion specifics, baseline list, and hyperparameters) are not reported in the text supplied here.

Why This Matters

Impact on research: The paper reframes mixer analysis away from deposit–withdrawal deanonymization toward a different question: which transactions are laundering-related, and which belong to the same case. It supplies a public, labeled benchmark where none existed, plus evidence that pairwise deanonymization links are neither necessary nor sufficient for AML investigation.

Real-world applications:

  • Exchange and bridge compliance teams screening incoming funds that may have passed through a mixer.
  • Chain-analytics vendors adding transaction-level mixer-laundering risk signals and case clustering to existing address-level tools.
  • Investigators reconstructing a single attack campaign split across many addresses, time windows, and laundering stages.
  • Mixer operators and block explorers, as shown by the paper's disclosure of four validated January 2026 incidents worth over $15 million to Etherscan and Tornado Cash.

Industry relevance: The findings target concrete compliance weaknesses. Tornado Cash performs no pool-contract-level risk screening, and Railgun's blacklist-based Private Proof of Innocence checks only deposits, is not retroactive, and lags on updates – falling short of the FATF Travel Rule's expectation that originator and beneficiary be assessed under the same standard. The reported performance suggests detection can be done from publicly observable on-chain behavior without breaking mixer privacy.

Future Directions

  • Extending beyond two mixers. The study covers Tornado Cash and Railgun, which accounted for about 95% of Ethereum mixer transactions in 2025; whether the five strategies and MixGuard transfer to other privacy mechanisms, chains, or cross-chain hops is untested.
  • Closing the bidirectional screening gap. The identified asymmetry between deposit-side and withdrawal-side compliance raises the question of whether withdrawal-side risk checks, retroactive re-screening, and faster blacklist updates are technically and politically feasible for mixer designs.
  • Handling label scarcity and case drift. With 27 cases, coverage of laundering behaviors and causes (including private key compromise, social engineering, smart contract exploits, logic flaws, and API key abuse) is finite; how well the model holds as attackers adapt to published strategies is an open question.
  • Operating under adversarial evasion and privacy constraints. The paper frames detection as achievable without breaking mixer privacy, but the boundary between legitimate privacy use and laundering – and how detection degrades if launderers deliberately imitate background behavior or cross the weak-tie structure – remains to be explored.

Target Audience

Blockchain security and AML researchers, on-chain analytics engineers, and compliance staff at exchanges, bridges, and custodians who need to reason about funds that have passed through privacy mixers. It is also useful for investigators and regulators seeking an empirical, case-level picture of how mixer laundering actually behaves, and for privacy researchers interested in where legitimate mixer use ends and abuse begins.

Authors’ abstract

Mixers protect privacy by concealing deposit--withdrawal links, but are also abused to launder illicit funds. Existing anti-money laundering studies do not specifically target mixer laundering, while mixer research focuses on deanonymization rather than identifying laundering-related transactions. Public reports remain fragmented, leaving no public case-level dataset for systematic measurement and detection. To fill this void, this paper presents the first comprehensive…

The full abstract on arXiv
Read the original paper