Research
StableAML: Machine Learning for Behavioral Wallet Detection in Stablecoin Anti-Money Laundering on Ethereum
Stablecoins such as USDT and USDC have become the preferred rail for moving illicit crypto funds, yet most AML detection research targets native assets like Bitcoin or Ether rather than stablecoin…

- arXiv
- 2602.17842
- Published
- 2026-02-19
- Authors
- Luciano Juvinski, Haochen Li, Alessio Brini
AI summary
Overview
Research area: Blockchain anti-money laundering (AML), applied machine learning, and stablecoin compliance analytics (cs.CR).
Technical level: Intermediate. The paper assumes familiarity with supervised classification, graph neural networks, and Ethereum token mechanics, but its argument is accessible to compliance and policy readers.
Scope: The paper builds StableAML, a labeled Ethereum dataset of USDT and USDC wallet behavior, and benchmarks tree-ensemble classifiers against graph neural networks for triaging suspicious wallets under realistic conditions of transaction-graph fragmentation.
What This Paper Is About
Stablecoins such as USDT and USDC have become the preferred rail for moving illicit crypto funds, yet most AML detection research targets native assets like Bitcoin or Ether rather than stablecoin transfer behavior. The authors ask whether a feature-engineered, domain-informed machine learning model can classify wallets as Normal, Cybercrime, or Blocklisted using only stablecoin transfer events, and whether such a model can outperform graph-based approaches when the transaction network is deliberately sparse and fragmented. The goal is a high-precision triage tool for compliance officers, not a legal determination of money laundering.
Key Contributions
-
StableAML dataset construction. The authors assemble what they describe as the first large-scale labeled dataset focused exclusively on stablecoins, drawing labels from Etherscan reports, disclosures from blockchain security firms (SlowMist, PeckShield), and the OFAC Specially Designated Nationals list, then joining them to the complete USDT and USDC transfer history on Ethereum between 2017-11-28 and 2025-08-08.
-
A 68-feature behavioral representation. Features are organized into four categories: Interaction Features (direct wallet–protocol relationships such as DeFi and centralized exchanges), Derived Network Features (multi-hop fund propagation and indirect exposure), Transfer-based Features (volume thresholds at $1,000, $5,000 and $10,000), and Temporal and Direct Features (timing patterns and intrinsic account properties).
-
Evaluation under a "Compliance Tunnel" constraint. Rather than testing graph models on an idealized, fully observed network, the authors restrict the graph to stablecoin edges over labeled wallets, producing edge density below 0.01, and compare tree ensembles against graph neural networks in that sparse regime.
-
Interpretability beyond binary detection. The framework is used to distinguish behavioral typologies, separating the high-velocity dispersion characteristic of cybercrime syndicates from the constrained, static footprints of sanctioned entities, with the stated aim of informing compliance under MiCA and the GENIUS Act while reducing unjustified asset freezes.
Main Findings
-
Tree ensembles outperform graph neural networks. Domain-informed tree ensemble models achieve a higher Macro-F1 score than the graph-based alternatives, which the authors attribute to the increasing fragmentation of stablecoin transaction networks. The exact Macro-F1, AUROC, and per-model scores are not reported in the available text; the results section is truncated.
-
Graph sparsity is structural, not an artifact. The graph is built over labeled wallets with edge density below 0.01. The authors frame this as a realistic property of the regulatory environment rather than a curation flaw, and interpret the model comparison as evidence of relative resilience to topological fragmentation rather than a general claim about architecture families.
-
Illicit wallets are structurally embedded in high-volume flows. In exploratory analysis, the
2ndWithOver10kfeature (second-degree connections to transfers above $10k) is zero-exposure for legitimate users but shows a distinct distribution for Blocklisted and Cybercrime wallets. -
Protocol preferences diverge sharply. Legitimate users favor centralized exchanges (
sentToCex), whereas illicit actors predominantly route funds through decentralized swap protocols. -
Distinct typologies are separable. The model differentiates cybercrime's complex, high-velocity dispersion from the constrained, static footprints of sanctioned entities, supporting a typology-aware rather than purely binary view of risk.
-
The dataset is heavily imbalanced at source and rebalanced by sampling. Of 54,998,597 unique wallets in the full transfer record, 363,887 were identified as malicious or suspicious. The two illicit categories account for less than 1% of active addresses at the macro-network level, so stratified sampling anchored on the Cybercrime class produced a 16,433-wallet modeling set with approximately 48.7% Normal, 36.5% Cybercrime, and 14.8% Blocklisted, split 80/20 into training and held-out test.
-
Illicit layering is fast and multi-hop. A worked example traces roughly 90 USDT withdrawn from Tornado Cash, 82.416 USDT forwarded to a second intermediary twenty minutes later, and a deposit into a Binance wallet less than 30 minutes after the initial withdrawal. A separate example shows four identical 50,000 USDT transfers totaling 200,000 USDT within roughly fifteen minutes.
Methodology in Plain English
The authors start from the full historical record of USDT and USDC transfers on Ethereum, which they report as 334,754,008 transactions across 54,998,597 unique wallets. They read the tokens' Transfer events rather than raw transactions, because event logs identify the true sender and recipient even when a relayer submits the transaction on someone else's behalf.
They label wallets into three groups. Normal wallets carry no negative flag from Chainalysis, TRM Labs, or Crystal Intelligence, and are additionally screened with a "Proximity Exclusion Filter" that removes any candidate within a 2- to 3-hop transaction radius of a known illicit node, so undetected accomplices do not pollute the benign class. Cybercrime wallets are direct perpetrators of exploits, hacks, scams, and phishing, sourced from public threat-intelligence reports. Blocklisted wallets are addresses sanctioned by OFAC or frozen at the smart-contract level, extracted by decoding the addBlocklist events of the USDT and USDC contracts.
All 68 behavioral features are computed at the full network level before any sampling, so that multi-hop metrics reflect each wallet's true position in the ecosystem rather than a truncated subgraph. The labeled sample is then drawn by stratified random sampling rather than synthetic oversampling, to preserve authentic transactional signal. Models are trained on 80% of wallets and evaluated on a disjoint 20%, using Macro-F1 as the primary metric because it weights each class equally under imbalance, with AUROC reported as secondary. The graph variant models wallets as nodes and aggregate stablecoin volume as directed, weighted edges, and performs static multiclass node classification with the same 68-dimensional feature vector per node.
Why This Matters
Research impact. The paper argues that existing AML research is built on native cryptoassets or on continuous graph topologies, leaving a gap for stablecoin-specific, feature-centric detection. It also proposes that as zero-knowledge privacy layers obscure native chains, stablecoin ledgers become the persistent auditable surface, which reframes where AML modeling effort should go.
Real-world applications:
-
Stablecoin issuer compliance. Tether and Circle have frozen over $4 billion in USDT and more than $1 billion in USDC. A wallet-level risk score can help prioritize which addresses to review before freezing, addressing the paper's concern about unjustified asset freezes.
-
Exchange and VASP onboarding. The proximity and typology signals support know-your-transaction screening at deposit time, particularly distinguishing exchange-bound legitimately behaving wallets from DeFi-swap-routing ones.
-
Law enforcement triage. The framework is positioned as a triage mechanism that ranks behavioral risk for investigators, who must then establish predicate offenses with off-chain context.
-
Regulatory audit under MiCA and the GENIUS Act. Both regimes impose AML obligations and auditability requirements on issuers; the paper frames its outputs as inputs to those verification workflows.
Industry relevance. The paper cites compliance costs of $136.5 billion annually in Europe and rule-based false positive rates above 95%, against interception of only 0.1% of global criminal funds. It also situates the work against the backdrop of $51 billion laundered through crypto in 2024 (about 0.14% of $9 trillion in on-chain transactions) and stablecoins accounting for over 84% of verified crypto fraud volumes in 2025.
Future Directions
-
Moving from static to real-time scoring. The authors explicitly describe the current work as static classification of behavioral signatures and a precursor to future real-time monitoring systems.
-
Extending beyond Ethereum USDT and USDC. The framework is scoped to two tokens on one chain, while the regulatory discussion spans Singapore's Payment Services Act, Hong Kong's Stablecoin Issuer Regulatory Regime, and China's digital currency framework.
-
Operating under genuine privacy layers. The paper tests a simulated fragmented graph, but the harder case is one where native chains are actually obscured by zero-knowledge shielding protocols such as Railgun; how much signal survives is unresolved.
-
Closing the gap between risk scoring and legal identification. The authors note that strict identification of money laundering requires off-chain context and proof of predicate offenses, and treat that gap as open.
Target Audience
Compliance and AML analysts at stablecoin issuers, exchanges, and virtual asset service providers; blockchain analytics researchers and practitioners; machine learning researchers working on graph-based financial crime detection; and policy and regulatory staff interpreting MiCA and GENIUS Act AML obligations who need a grounded view of what on-chain behavioral models can and cannot deliver.
Authors’ abstract
Global illicit fund flows exceed an estimated…
The full abstract on arXiv