Skip to content
Crypto.world

Research

ART: A Graph-based Framework for Investigating Illicit Activity in Monero via Address-Ring-Transaction Structures

Monero's RingCT protocol hides the true spender among decoy inputs, conceals amounts, and routes funds through one-time stealth addresses, which makes the address-based tracing used on transparent…

ART: A Graph-based Framework for Investigating Illicit Activity in Monero via Address-Ring-Transaction Structures
arXiv
2511.16192
Published
2025-11-20
Authors
Andrea Venturi, Imanol Jerico-Yoldi, Francesco Zola, Raul Orduna

AI summary

Overview

Research area: Blockchain forensics and privacy-preserving cryptocurrency analysis — specifically behavioral detection of illicit activity on Monero using graph-based representations and machine learning.

Technical level: Intermediate. Readers should be comfortable with basic graph concepts (nodes, edges, hops) and standard classification metrics (precision, recall, F1), but no deep cryptographic background is required.

Scope: A small-scale case study that builds Address-Ring-Transaction (ART) graphs from 19 known Lazarus Group / WannaCry 2.0 Monero transactions plus 150 sampled licit transactions, extracts structural and temporal features, and trains a Random Forest classifier to test whether behavioral detection in Monero is feasible at all.

What This Paper Is About

Monero's RingCT protocol hides the true spender among decoy inputs, conceals amounts, and routes funds through one-time stealth addresses, which makes the address-based tracing used on transparent chains like Bitcoin largely ineffective. Rather than attacking the cryptography or trying to de-anonymize individual inputs, the authors ask a different question: can the shape and timing of a transaction's neighborhood reveal a criminal operating pattern that a machine learning model can recognize elsewhere?

The goal is a first, deliberately partial step toward investigator-facing tooling — a method that learns a "modus operandi" signature from already-flagged transactions and flags other transactions that look behaviorally similar.

Key Contributions

  1. The Address-Ring-Transaction (ART) graph. A new graph representation tailored to Monero's ring-signature protocol, with three node types — transactions, addresses, and rings — that captures the indirect links created when the same stealth address appears in multiple rings. It is an adaptation of the conventional address-transaction graph used in blockchain analytics, redesigned because Monero has no persistent user addresses to anchor behavior on.

  2. A hierarchical (n-hop) expansion scheme with a feature taxonomy. The graph is expanded in the output direction hop by hop, and features are organized into 0-hop features (taken from the seed transaction itself) and i-hop features (statistics over transactions at the i-th hop). This gives a tunable way to widen the structural and temporal context around a transaction, at exponentially increasing computational cost.

  3. A behavioral classification pipeline. A four-phase method — data acquisition, ART-graph construction, feature extraction, classification — that trains binary classifiers and evaluates them with Precision, Recall, and F1-score, treating flagged transactions as the positive class.

  4. First empirical evidence that behavioral detection in Monero is feasible. To the authors' knowledge this is the first application of behavioral graph analysis to Monero; prior work on Monero focused on traceability heuristics and protocol vulnerabilities, and prior behavioral graph work focused on transparent cryptocurrencies.

Main Findings

  • Classifier performance: The Random Forest model reached Precision 1.000, Recall 0.750, and F1-score 0.857 on the test set. The authors describe these as encouraging given the limited dataset size.

  • Perfect precision means zero false positives on the test set: every transaction the model labeled illicit was in fact labeled illicit in the ground truth. The authors note this indicates the ART-graph features capture structure and behavior associated with the Lazarus Group activity.

  • The recall gap signals missed detections. A recall of 0.750 means false negatives occurred — some illicit transactions went undetected. The authors attribute this partly to the Lazarus Group's heterogeneous and deliberately obfuscated behavior, which can push individual transactions outside the dominant profile the model learned.

  • Dataset composition: 19 Monero transactions tied to the WannaCry 2.0 campaign (reported as linked to the Lazarus Group), split by the source report into 8 entering transactions (all on 03 August 2017), 3 consolidating transactions (17 August 2017), and 8 exiting transactions (02 November 2017), all labeled positive. 150 additional transactions randomly sampled from the same three-month period served as negatives. RingCT was already active on Monero when these operations occurred.

  • Class imbalance handling: SMOTE (Synthetic Minority Oversampling Technique) was applied to offset the 19-versus-150 imbalance by generating synthetic minority-class samples. The paper states the dataset was divided 80 (the text is truncated at this point, so the exact train/test ratio is not fully reported in the available content).

  • Graph depth used: A 2-hop ART-graph was constructed for every transaction in the dataset, expanded by identifying output addresses, connecting them to the rings they participate in (as real spender or as mixin), and linking those rings back to their transactions.

  • Acknowledged constraints: The authors flag the very limited support in both classes, the scarcity of publicly available labeled Monero data, overfitting risk to the Lazarus Group's specific pattern during the analyzed window, the narrow single-campaign and narrow-time-window scope limiting generalizability, the random negative sampling possibly missing other modi operandi, and SMOTE generating interpolations rather than genuine diversity.

Methodology in Plain English

The researchers started from a practical reality: on Monero you cannot follow an address, because recipients get fresh one-time addresses and every input hides behind decoys. So they made the transaction the unit of analysis instead.

For each transaction, they draw a small graph. The transaction is a node. On its input side sit "rings" — the groups of decoy addresses that hide the real spender. On its output side sit one-time addresses. Those addresses get reused as ring members in later transactions, which quietly stitches otherwise unrelated transactions together. That stitched structure is the ART graph.

They then expanded each graph outward two hops, adding the transactions that produced the addresses and rings at each layer. From this picture they computed two kinds of numbers. Structural ones describe the transaction's shape: how many rings, how big the rings are, how many outputs, the fee, how many distinct ring members appear across all inputs. Temporal ones describe timing: the gaps between the seed transaction and the transactions that created its ring members, summarized as minimum, maximum, mean, standard deviation, and median, repeated at each hop for the neighboring transactions' ring counts and ring sizes too.

Those numbers became the feature vector for a Random Forest classifier. Since only 19 positive examples existed against 150 negatives, they used SMOTE to synthesize extra positive examples so the model would not simply learn to say "no." The model was then scored on held-out data using precision, recall, and F1 — chosen because in this setting a false accusation (false positive) and a missed illicit transaction (false negative) carry different costs.

Why This Matters

Impact on research: The paper reframes Monero forensics as a behavioral problem rather than a cryptographic one. Most prior Monero literature tries to defeat ring signatures or find protocol weaknesses — approaches that worked against early protocol versions and are no longer applicable, or that need stringent conditions and heavy computation. This work opens a complementary track: even if you cannot identify the spender, you may still recognize the operating pattern. It also demonstrates the ART graph as a reusable representation for a chain where the classic address-transaction graph simply does not apply.

Real-world applications:

  • Ransomware response: Tracing and characterizing how ransom proceeds are converted through Monero, as in the WannaCry 2.0 / Lazarus Group and LockBit cases cited in the paper.
  • Darknet market and CSAM vendor investigations: The paper cites a January 2024 Chainalysis report on CSAM vendors shifting toward instant exchanges offering Monero, a context where pattern detection could support prioritization.
  • Money laundering detection: Flagging conversion and consolidation flows that resemble known laundering routines rather than known addresses.
  • Investigative triage: Ranking or filtering blockchain transactions for analyst review when exhaustive manual examination is impossible under privacy guarantees.

Industry relevance: For blockchain analytics vendors and exchange compliance teams, the work suggests a feature family and graph schema that could complement existing transparent-chain tooling for privacy coins — an area where commercial coverage is thin precisely because address-based heuristics do not transfer.

Future Directions

  • Expanding the labeled dataset, which the authors identify as the key limitation: both positive and negative classes are too small, and publicly available labeled Monero data is extremely scarce. Larger data would test whether the recall gap closes and reduce overfitting to one campaign.
  • More precise selection of negative (licit) transactions, replacing random sampling with a strategy that accounts for other modi operandi — including ones that might superficially resemble the positive class.
  • Testing generalization beyond a single campaign and time window, since the current model is trained on one campaign inside a narrow three-month period and its transferability to other illicit behaviors or future network activity is unknown.
  • Scaling the graph depth and validating the cost trade-off, since larger n widens the behavioral context but the ART graph grows exponentially per hop — leaving open which n is worth the computation in practice.

Target Audience

Law enforcement analysts and blockchain forensics practitioners working on privacy-preserving cryptocurrencies; academic researchers in cryptocurrency forensics, graph-based anomaly detection, and anti-money-laundering; and compliance or threat-intelligence teams at analytics vendors and exchanges who need to understand both the promise and the current scale limits of behavioral detection on Monero. Readers from adjacent fields such as graph machine learning will find the ART-graph representation and n-hop feature taxonomy the most transferable parts, while investigators should note that the study is a 19-transaction case study and not a deployable detector.

Authors’ abstract

As Law Enforcement Agencies advance in cryptocurrency forensics, criminal actors aiming to conceal illicit fund movements increasingly turn to "mixin" services or privacy-based cryptocurrencies. Monero stands out as a leading choice due to its strong privacy preserving and untraceability properties, making conventional blockchain analysis ineffective. Understanding the behavior and operational patterns of criminal actors within Monero is therefore challenging and it is…

The full abstract on arXiv
Read the original paper