Pricing Cross-Bank Context in Federated Anti-Money Laundering

Lidia Losavio; Dario Fenoglio; Martin Gjoreski; Marc Langheinrich

cs.CROct 26, 2026

Pricing Cross-Bank Context in Federated Anti-Money Laundering

Authors: Lidia Losavio, Dario Fenoglio, Martin Gjoreski, Marc Langheinrich

Organizations: Università della Svizzera italiana (USI), Lugano, Switzerland

Abstract

Most graph-based Anti-Money Laundering (AML) methods assume access to a pooled transaction graph containing all accounts and payments, which in reality does not exist. Federated learning offers the necessary privacy guarantees but it remains unclear how money laundering patterns spanning separate graphs can effectively be detected. We exploit transactions already recorded by both endpoint banks to approximate remote account states locally, enabling multi-layer Graph Neural Network (GNN) computation without additional account-specific exchange. Starting from this baseline, we test whether the remaining context is worth sharing, using a disclosure ladder that adds two transaction counts per remote account or the counterparty's own GNN states. On AMLworld HI-Small, across six GNN backbones and up to 32 silos, the observable baseline yields the largest gain and keeps it as fragmentation shrinks the visible slice; counts add a robust correction, while exact remote states help only in specific regimes. In controlled privacy--utility studies, the noise needed to protect shared states erases their detection value before it buys meaningful privacy, whereas counts stay useful even under strong privacy protection. These findings support a practical rule for banks: use what they already record, and formally protect the little that must move.

Figures & tables

Figure 1. Money laundering patterns ( Altman et al., 2023 ) . The blue node indicates the main laundering account.

Figure 2. Observable builds remote context locally. (a) $\mathcal{T}_{A}(b)$ here denotes the visible incoming slice. (b) Bank $A$ uses it to approximate $b$ ’s first-hop state under $\theta_{t}$ , without exchange. (c) The approximation is cached during local updates and refreshed at synchronization.

Figure 3. HI-Small fragmentation study: pooled test F1 score, mean

\pm

std over three seeds. Dashed black line and grey band: centralized reference; FedSage+ (pink) repeats across panels as a backbone-independent reference.

Figure 4. Protected first-hop states, controlled tasks.

\Delta

PR-AUC = private Eobi

-

Observable , paired 95% intervals. Horizontal axis: distance-scoped

\epsilon_{m}(r_{90})

\delta=10^{-5}

(not worst-case

\epsilon_{e}

); smaller is stronger. Distributional and target-selective residual panels show the utility difference between private state exchange and Observable. Selective exchange is useful only at the weaker end of the distance-scoped privacy range; distributional exchange has no confirmed noisy gain.

Figure 5. Protected total counts, cardinal task (PNA,

\rho=.1

\lambda=.5

\Delta

PR-AUC relative to Observable , paired 95% intervals; horizontal axis is pure event-level

\epsilon

per count-vector release. Private count utility remains above Observable for all tested privacy budgets, including epsilon 0.01, and approaches the clean count gain as epsilon increases.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6. Three prices for cross-bank context at a frontier account

b

: Observable sends no payload, Observable+degree two degree counts, and Eobi the owner-computed cut state (first-hop state shown; the benchmark record also carries

h_{b}^{(2)}

when consumed by the readout, i.e.,

2d

floats); raw residual transactions remain at

B

Figure 7. Per-typology interface value on HI-Small (TransConv) at $M{=}4$ and $M{=}32$ : paired recall difference vs. Observable , in percentage points; $n$ counts typology-annotated test positives.

Residual	Pooler	Observable	+count	Eobi
Distributional	GIN	$.853$	$.854$	$.855$
Cardinal	PNA	$.763$	$.796$	$.806$
Selective	TransConv	$.822$	$.824$	$.869$

Appendix

Table 1. Unprotected residual value at low coverage (

\rho=.1

\lambda=.75

): three-seed mean PR-AUC. Here +count releases one total count.

Figure 8. Controlled value of unprotected boundary-state exchange over the $(\lambda,\rho)$ grid. Each cell reports the three-seed mean $\Delta$ PR-AUC ( Eobi $-$ Observable ); columns vary the target-edge contribution $\lambda$ , rows the observable coverage $\rho$ . Dashed outlines mark $\rho=.3$ , within the HI-Small coverage range $.245$ – $.366$ ; the solid box marks the TransConv $\lambda=.75$ reference slice. Matching coverage alone does not equate a synthetic task with the benchmark. Heatmaps show the incremental PR-AUC of boundary-state exchange over Observable for three synthetic residual classes as observable coverage and the target-edge label contribution vary. Gains concentrate at lower coverage when residual context remains relevant.

Anti-Money Laundering Detection

See all 22 papers

SALT-GNN: Handling Dense Neighborhoods in Anti-Money Laundering Graphs via Statistics-Aware Attention
Jul 11, 2026Lidia Losavio, Francesco Sovrano, Dario Fenoglio +2Graph Attention Networks Graph Neural Networks
Università della Svizzera italiana (USI) Lugano, Switzerland
Privacy-Preserving GNN Inference on Financial Transaction Graphs via Homomorphic Encryption
Oct 13, 2026H. Çağrı Bilgi, Kemal Mutluergil, Ceren Yıldırım +2Anti-Money Laundering Detection Privacy-Preserving ML
Sabancı University, Istanbul, Türkiye
LineMVGNN: Anti-Money Laundering with Line-Graph-Assisted Multi-View Graph Neural Networks
Mar 24, 2026Chung-Hoo Poon, James Kwok, Calvin Chow +1Anti-Money Laundering Detection Financial Fraud Detection
Logistics and Supply Chain MultiTech R&D Centre, Level 11, Cyberport 2, 100 Cyberport Road, Hong Kong · Department of Computer Science and Engineering, The Hong Kong University of Science and Technology, Clear Water Bay, Kowloon, Hong Kong

Jul 11, 2026J/K move · Enter open · S save

SALT-GNN: Handling Dense Neighborhoods in Anti-Money Laundering Graphs via Statistics-Aware Attention

Lidia Losavio, Francesco Sovrano, Dario Fenoglio, Martin Gjoreski+1

Università della Svizzera italiana (USI) Lugano, Switzerland

Graph Attention Networks Graph Neural Networks Financial Fraud Detection Anti-Money Laundering Detection Anti-Money Laundering

Open paper AskPDF arXiv

Money laundering threatens financial stability and exposes institutions to penalties, motivating automated detection. Because laundering schemes often emerge through relational patterns, graph neural networks (GNNs) are increasingly used for anti-money laundering (AML). Yet AML GNNs are typically evaluated with aggregate metrics such as overall F1 score, which hide an operational issue: high-activity recipient accounts concentrate many incoming transactions, making suspicious signals harder to isolate and costlier to investigate. We introduce a recipient-degree stratified evaluation that reports standard AML metrics across recipient-context density. Across three datasets (HI-Small, HI-Medium, and AMLSim-32k-5%), it reveals consistent degradation in dense recipient contexts, which we trace to three GNN characteristics: two known limitations that AML amplifies, i.e., (1) multiset non-discriminability and (2) cardinality blindness, and (3) an attention-specific effect: in dense neighborhoods, normalized attention attenuates weak but pattern-relevant multi-hop signals. Guided by this diagnosis, we propose SALT-GNN, a lightweight statistics-aware architecture that fuses degree-aware statistical aggregation with attention at each message-passing layer, so distributional and cardinality information shapes the node states used by subsequent attention steps. Ablations support fusion placement as a key factor in dense-context performance. On HI-Small and HI-Medium, SALT-GNN uses up to 77% fewer parameters than task-specific graph-transformer baselines while improving dense-context F1 score by 3-6 points; on AMLSim-32k-5%, it improves highest-degree F1 score by 16-20 points. The gains hold for both Transformer- and GAT-style attention, indicating that the benefit comes from where statistical and attentional evidence is fused rather than from a specific attention operator.

Abstract

Anti-Money Laundering Detection

Explore similar work

SALT-GNN: Handling Dense Neighborhoods in Anti-Money Laundering Graphs via Statistics-Aware Attention

Privacy-Preserving GNN Inference on Financial Transaction Graphs via Homomorphic Encryption

LineMVGNN: Anti-Money Laundering with Line-Graph-Assisted Multi-View Graph Neural Networks