EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Subgraph-Level Temporal Graph Transformer With Laundering-Motif Prototypes for Anti-Money Laundering on Blockchains

Yizhen Lin, Xinran Yue, Menglin Bian, Rufeng Zhu, Dai Teng

IEEE Access · 2026 · 14 · 139135-139159

DOI: 10.1109/access.2026.3731160 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Yizhen Lin und 4 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 70 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Money laundering on a public blockchain poses a representation-learning problem that current models leave unsolved: the discriminative signal is a subgraph whose meaning lies jointly in topology and in time, and the resulting alert must be auditable. Message-passing networks are provably unable to count the chain and cycle substructures that define laundering motifs and over-squash long chains; static graph transformers remove that bottleneck but discard edge timing, a first-class signal (fund velocity, dormant-then-active mules); and continuous-time temporal graph networks encode timing yet emit only node-level embeddings, never a single time-aware representation of a laundering pattern. We formalize this gap and propose Stemp, a Subgraph-level TEmporal graph transformer with Motif Prototypes, whose contributions are methodological. First, we introduce an attention bias that is jointly structural and temporal, shifting logits by a learnable shortest-path term and by a learnable function of the inter-node time gap; we prove that this is strictly more expressive than any static or topology-only-biased counterpart, and that it reduces the depth needed to couple the ends of a length- $k$ chain from $\Omega (k)$ to one layer. Second, we lift case-based prototype reasoning to sets of node tokens, so a motif is matched even when it occupies part of a candidate subgraph, and the decision logit decomposes additively over motif activations, giving an auditable attribution. Third, we transfer a Bitcoin encoder pre-trained on node-level Elliptic1 across prediction levels to subgraph-level Elliptic2 and across chains to Ethereum, with an explicit leakage audit. Stemp improves minority-class $F_{1}$ by 4.8 points over the strongest of sixteen baselines on Elliptic2 ( $0.722\!\rightarrow \!0.770$ ; exact paired Wilcoxon $p{=}0.002$ over ten seeds, bootstrap and McNemar corroboration, multiplicity control), by 6.9 points at 10% label supervision and by 5.9 points cross-chain. Label-noise, adaptive-adversary, scalability and end-to-end deployment studies, and quantitative validation of prototype–motif alignment (purity 0.83, injected-motif top-1 recovery 0.88), delimit where the model works and where it does not.

Zitieren

APA

Lin, Y., Yue, X., Bian, M., Zhu, R., & Teng, D. (2026). Subgraph-Level Temporal Graph Transformer With Laundering-Motif Prototypes for Anti-Money Laundering on Blockchains. IEEE Access, 14, 139135-139159. https://doi.org/10.1109/access.2026.3731160

Vancouver

Lin Y, Yue X, Bian M, Zhu R, Teng D. Subgraph-Level Temporal Graph Transformer With Laundering-Motif Prototypes for Anti-Money Laundering on Blockchains. IEEE Access. 2026;14:139135-139159. doi:10.1109/access.2026.3731160

BibTeX
@article{Lin2026,
  author = {Lin, Yizhen and Yue, Xinran and Bian, Menglin and Zhu, Rufeng and Teng, Dai},
  title = {Subgraph-Level Temporal Graph Transformer With Laundering-Motif Prototypes for Anti-Money Laundering on Blockchains},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {139135--139159},
  doi = {10.1109/access.2026.3731160}
}
Kostenlos starten →