Research
Population-Calibrated Graph Screening at 835-Million-Address Scale, with Label-Free Transfer to New Chains
Compliance screening of crypto addresses today is mostly a lookup against sanctions lists plus clustering heuristics, which says nothing useful about addresses that are not on a list and breaks down…

- arXiv
- 2609.03036
- Published
- 2026-09-02
- Authors
- Yury Korolev
AI summary
Overview
Research area: Blockchain forensics and compliance screening, specifically graph representation learning applied to anti-money-laundering and illicit-address detection across EVM chains.
Technical level: Advanced. The paper assumes familiarity with graph neural networks (GraphSAGE), inductive node embeddings, calibration of classifier scores, reinforcement learning (PPO), and production ML serving concerns.
Scope: A deployed five-chain EVM screening system built on a single 835,330,427-address transaction graph, covering its substrate, population-level threshold calibration, label-free transfer to chains with no training labels, retrospective lead-time evaluation against external registry events, serving path, and a reinforcement-learning adversarial test harness.
What This Paper Is About
Compliance screening of crypto addresses today is mostly a lookup against sanctions lists plus clustering heuristics, which says nothing useful about addresses that are not on a list and breaks down entirely on chains with no label coverage. This paper describes a production system that instead scores an address by where it sits in a large multi-chain transaction graph, and it shows that the resulting model can be applied to new chains that contributed no labels to training. The paper's secondary aim is evaluation discipline: it reports where its own numbers should not be trusted, including a measured blind spot against synthetically generated adversary behaviour.
Key Contributions
-
A single multi-chain graph substrate. 835,330,427 addresses with 25 features each and 15,826,261,934 edges, spanning Ethereum, Tron, Base, Arbitrum and Gnosis in one node space with contiguous per-chain index segments (Ethereum 323,139,851; Tron 331,982,925; Base 131,409,131; Arbitrum 39,134,190; Gnosis 9,664,330), so that the screening signal lives in an inductive graph representation rather than in target-chain labels.
-
Calibration by exact full-population scan. Decision thresholds are exact quantiles of the score distribution measured by scanning every score in the graph, per chain segment, so the alert volume is known in advance. The paper publishes the complete threshold grid with realised flagged counts and documents two measured pathologies that make nominal alert-rate points non-comparable across chains.
-
A label-free transfer result under an explicit per-component protocol. Heads trained with no Base, Arbitrum or Gnosis labels recall 0.8598 / 0.8182 / 0.9967 of held-out positives on those chains at a 10^-3 population alert rate, with a stated accounting of which components did and did not consume target-chain data.
-
A lead-time replay against 68 external registry events, with a stated static-replay convention and an event-level random baseline (40 of 68 detected, 58.8 percent, a ×152 lift, median first on-chain appearance 528.8 days on Ethereum and 647.8 days on Tron before public designation), plus a production serving path gated by bit-exactness and a 2,882-address drift panel.
Main Findings
-
Label-free transfer works at a 0.1 percent population alert budget. Heads trained with no labels from the target chains recalled 0.8598 (Base), 0.8182 (Arbitrum) and 0.9967 (Gnosis) of held-out positives. The authors flag the Gnosis figure as the least informative of the three, because its threshold (0.5431) is the lowest in the grid and its score distribution is compressed toward zero.
-
Nominal alert rates do not realise equally across chains. The Gnosis threshold at 10^-2 is 0.0486, far below every other chain, because that chain's score distribution is compressed toward zero. The Base flagged count at 10^-4 is 281 where a strict 10^-4 of 131 million would be roughly 13,141, because the quantile lands on a plateau of tied scores.
-
The population alert rate is a budget, not a false-positive rate. Because the population contains an unknown number of true positives, the false share of the flagged 0.1 percent is not identified; the paper reserves "FPR" for pools with known negatives.
-
External-ground-truth replay. At the fincrime-head 10^-3 threshold, 23 of 42 Ethereum events were detected (54.8 percent, ×139 lift, median lead 528.8 days) and 17 of 26 Tron events (65.4 percent, ×174 lift, median lead 647.8 days); the EVM total is 40 of 68 (58.8 percent, ×152, median lead 578.4 days). The event-level null expects 0.26 of 68 detected, against 40 observed. On the audited-date subset (58 of 68 events, after 28 registry default dates were replaced with the publishing OFAC action), 33 were detected at a median lead of 647.8 days. 37 of the 40 detections precede publication.
-
The replay is explicitly static. Features are aggregates over each address's complete history, including post-first-appearance activity, so the replay establishes a retrospective association, not an early warning. The time-censored replay (features truncated at time T) has not been run.
-
Comparison to the previous production line. Under identical protocol, splits, seed and folds, five of five chains improved: Ethereum 0.7607 to 0.8314 (+0.0707), Tron 0.5655 to 0.5783 (+0.0128), Base 0.8518 to 0.8598 (+0.0080), Arbitrum 0.7818 to 0.8182 (+0.0364), Gnosis 0.9934 to 0.9967 (+0.0033).
-
Two heads beat one. A single head over a union pool measured 0.44 against 0.596 for either-head on the same checkpoint, with 42 percent of the union pool being the other head's nodes. On Ethereum at 10^-3, either-head recall (0.8427) is below the fincrime head alone on the fincrime subset (0.8537) because the union adds 10,007 agent-class positives whose head is weaker at that threshold.
-
Labels have a modest marginal return. Doubling the labelled pool buys +0.018 recall on Ethereum and +0.014 on Tron across 5 folds, which is the empirical argument for putting the signal in the graph rather than in the list. A controlled feature-only baseline reaches 0.837 leak-free test recall at 10^-3 against the graph's 0.9289, and falls 0.2 to 0.3 behind on Base and Arbitrum.
-
Tron is the weakest chain and the ceiling is label coverage. Either-head recall 0.5783 at 10^-3 against 0.8314 on Ethereum, on a pool of only 1,482 positives (σ 0.048 across folds, itself a small-sample artefact). Gradient boosting matches the graph on Tron, supporting the label-coverage reading over an architecture explanation.
-
Serving is bit-exact and fast. For a well-known mixer contract the deployed path returns 0.8933565020561218, bit-identical to the offline computation, under model version 63.v75-5chain-d465968d17fde8ce, which is pinned in every response. End-to-end authenticated HTTPS latency was p50 151 ms and p95 256 ms (n=15, one warm address); server-side over n=5,000 calls cycling 144 registry addresses, p50 175.9 / p95 188.1 / p99 922.5 ms, with first touch of an address at p50 636.7 ms and a minimum observed 0.9 ms, the spread being chunk-cache effects.
-
The adversarial harness finds a real blind spot, and the snapshot protocol decides the verdict. On a detector-stratified snapshot the baseline mean was 0.2729 (p95 0.8635) and all eight archetypes read as detected; on a uniformly random snapshot (n=2,000, in-graph coverage 0.8894, baseline mean 0.1092, p95 0.4257) all eight read as blind spots, with the largest elevation Δ = +0.0366, below the pre-fixed +0.05 threshold. The authors conclude that on a snapshot drawn independently of the detector, the deployed heads do not separate generated activity from background.
-
An ablation contradicted the authors' own hypothesis. The expected cause of policy collapse was uninformative observations; an ablation with structural embeddings zeroed still passes 7/8 audit criteria, so the recurrence, not the embeddings, is what fixed the collapse.
-
The agent class audit is clean to within 8 addresses. All 10,007 agent-class positives in the Ethereum pool are explained by a label row, with 0 behavioural-only positives. The yield-contract hypothesis that originally motivated the class contributed 0 rows to training; only 8 specific addresses reached training through an older store epoch, a 0.0008 (0.08 percent) share of the pool's agent positives. The largest actual contributors are protocol-level tags: DEX routers, MEV bots, AVS operators, vault strategies.
-
Cross-chain entity resolution failed. Cluster purity was 0.410 on Ethereum and 0.333 on Tron against a 0.90 acceptance floor, with 0 of 168 candidate pairs resolved cross-chain. The cause is structural: the graph is block-diagonal by construction (726 cross-chain edges out of 10,728,037,368 in the earlier two-chain merge; chain centroids at cosine 0.24; cross-chain retrieval 7.2 percent at rank 100; the current five-chain merge has 953 cross-chain links).
-
Self-supervision lost. The self-supervised recipe scored 0.559–0.570 against 0.642, closing off a "just pretrain on the unlabelled graph" reading of the label-sparsity argument.
Methodology in Plain English
The authors build one large graph in which every address across five chains is a node and every transaction relationship is an edge. Each node carries 25 structural and behavioural summaries, such as transaction counts, value statistics, temporal spread, counterparty diversity and contract-interaction ratios. A three-layer GraphSAGE encoder with mean aggregation (width 128, dropout 0.3, per-chain LayerNorm on the output embedding) turns each node into a 128-dimensional vector once, offline, and the vectors are stored so that a new address is scored by lookup rather than by retraining.
Two small MLP heads sit on that shared embedding: one for financial crime, one for autonomous agents. Training warm-starts from the previous backbone, uses Adam at 5·10^-4 for the backbone and 10^-3 for the heads, batch 2,048, two-hop neighbour sampling with fanout [25, 10], 3 epochs and seed 42, with splits held verbatim from the previous iteration so comparisons are like-for-like. The validation trajectory across the three epochs was 0.9088 to 0.9220 to 0.9303, with no early stopping. Training took 81.1–82.3 minutes on one 8×A100 node; full-graph embedding emission took 122.4–124.7 minutes on a second identical node, producing 1,671 chunks of 500K rows at fp16, 213,848,356,544 bytes, uploaded in 11 minutes.
Rather than picking a threshold to optimise a validation metric, they scan every score in the graph, per chain segment, and set thresholds at exact population quantiles. For transfer, they hold out 20 percent of each target chain's labelled positives, sample 500,000 negatives per chain, and apply the target chain's own population threshold, so no target-chain labels enter head training while target-chain structure, normalisation statistics and quantiles do.
For the lead-time study they take 68 external registry events (designations, hacks, exchange collapses) with addresses on Ethereum and Tron, flag any address whose score meets its chain's 10^-3 threshold, and compare that address's first on-chain appearance, taken from the substrate itself, against the event's public publication date. Because each event carries 1 to 9 in-graph addresses (median 4), the random baseline is computed at event level as 1-(1-10^-3)^m.
Finally, they train eight recurrent adversary archetypes with PPO in a simulator built from real chain snapshots, audit those policies for reward hacking and degeneracy against 8 criteria, and score the generated activity against two differently constructed evaluation snapshots to test whether the evaluation protocol itself is trustworthy.
Why This Matters
Impact on research. The paper's methodological contributions are as important as its performance numbers. Publicly recording realised flagged counts next to nominal alert rates, stating a static-replay convention instead of calling it a backtest, publishing the leakage control across every training pool that contributes positives, and showing that a detector-stratified evaluation snapshot inverts the verdict for all eight adversary archetypes are practices the applied security-measurement literature has largely catalogued as pitfalls without demonstrating at this scale. The paper also reports a failed cross-chain entity-resolution attempt and a losing self-supervised recipe, both usually absent from published results.
Real-world applications:
- Compliance triage at exchanges and custodians, where an alert budget that can be stated in advance (for example 0.1 percent of a chain's addresses) is what an operations team can actually staff.
- Onboarding a newly supported chain onto a screening surface before any curated labels exist for it, avoiding a months-to-years label-acquisition lag.
- Prospective risk review of addresses that have been on chain for a long time but were never publicly designated, using the retrospective lead-time evidence as a prioritisation signal (with the static-replay caveat attached).
- Adversary simulation for control testing, where red teams need generated behaviour to probe whether a detector generalises beyond its label distribution.
Industry relevance. The open/closed split is deliberately drawn to be reusable. The released tooling covers the label-production schema and parsers, registry scrapers for nine jurisdictions, VASP, mixer and bridge registry builders, the tag-pack generator, the adversarial simulator (26 modules, 13,064 lines, 430 tests) and the detector training and evaluation harness, across four repositories. Withheld are the trained weights, the substrate, the emitted embeddings, the address and provenance indices and the serving stack. The +0.018 figure quantifies why: the tooling alone does not reconstitute the system.
Future Directions
-
A time-censored replay. The measurement the paper says it would take to claim early warning: truncate each address's history at time T, rebuild features and embedding at T, and score at T. The current 528.8-day Ethereum median is explicitly not that claim.
-
Repeating the dead-column census on the 25-feature matrix. On an earlier 20-feature epoch, 7 features were identically zero on Base, Arbitrum and Gnosis; the census has not been repeated on the current line, so no effective-dimensionality claim is carried for the transfer results.
-
Full-graph community structure. The 506M-node clustering pass has never completed, because the index build exceeded single-device memory (a 259 GB payload against 80 GB) and multi-device sharding is not implemented, leaving community-level statements local rather than global.
-
Closing the label gap on Tron and building agent-class pools off Ethereum. All 10,007 agent-class evaluation positives are Ethereum, so every ai_agent figure in the paper is an Ethereum figure; the Tron gap is attributed to label coverage rather than architecture, which makes label acquisition there the natural next step. Co-training the detector against the simulator is also named but explicitly placed outside the claim set.
Target Audience
Applied researchers and engineers working on graph learning for financial crime, blockchain analytics teams at exchanges, custodians and compliance vendors, and ML platform engineers who need to reason about population-level calibration, bit-exact serving and evaluation-snapshot confounds. Reviewers and auditors of deployed detection systems will find the limitations section and the open/closed boundary more useful than the headline recall figures. Readers wanting an introduction to blockchain forensics should look elsewhere; the paper assumes comfort with GNN training recipes and reinforcement-learning evaluation.
Authors’ abstract
Compliance screening of blockchain addresses is, in practice, a lookup against sanctions registries plus clustering heuristics; it fails on unlabelled addresses and on chains with no label coverage at all. We describe a deployed system that scores an address by its position in a multi-chain transaction graph rather than by its presence in a list. The substrate is a single graph of 835,330,427 addresses and 15,826,261,934 edges across five EVM chains; a shared inductive…
The full abstract on arXiv