acceptodds
Withdrawn from ICLR 2027

Periscope: A Single Bank's View into Cross-Institutional Laundering

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Graph neural networks have become the dominant approach to anti-money laundering (AML), yet progress is held back by the scarcity of public data that reflect how banking systems actually operate. Existing datasets typically compress banking records into account-to-account transfer graphs, discarding the ownership ties and cross-entity relations on which practical risk analysis relies. This mismatch restricts both the credibility of public evaluation and the development of models able to exploit richer financial structure. We present PERISCOPE, a simulation-based AML testbed grounded in real banking data. Unlike prior public datasets, PERISCOPE retains the complete customer, account, and transaction schema of a single-bank operational setting, including ownership relations over both internal and external accounts, together with fine-grained transaction-level timestamps. We document the construction pipeline, assess the fidelity of the simulation, and provide a detailed characterization of anomalous users and accounts. Our analysis shows that associations among external entities reveal additional structural links between suspicious multi-account users that remain invisible in typical AML scenarios. Building on PERISCOPE, we conduct a systematic empirical study at both the user and the account level. The results show that retaining the richer structure consistently improves AML performance over simplified transfer-graph constructions. Meanwhile, current leading specialized AML models bring only limited gains over generic graph baselines and scale poorly on our dense data. These findings indicate that the main bottleneck of public AML research lies not only in data scarcity but also in the overly simplified graph abstractions adopted by existing datasets. Overall, PERISCOPE offers a grounded single-bank AML testbed, together with concrete evidence on how customer, account, and transaction structure and external-entity associations advance graph-based AML evaluation and modeling.

Primary area:
datasets and benchmarks
Keywords:
Money laundering detection, Graph benchmark, Graph neural networks

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_fkvEOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces PERISCOPE, a simulation-based AML benchmark grounded in real banking data that preserves heterogeneous structure—internal users, internal accounts, external entities, and external accounts—along with ownership relations and transaction-level timestamps. Using a deep graph simulation framework from prior work, the authors construct a synthetic dataset, validate its fidelity against the original data, and conduct empirical studies showing that heterogeneous graph constructions consistently outperform simplified homogeneous abstractions for AML detection.

Critical strengths
  1. The analysis of anomalous versus normal accounts (degree, inter-transaction intervals, neighbor features) and the account-level vs. entity-level overlap study provide useful qualitative insights into how multi-account laundering behaviors manifest structurally.

2.The authors explicitly discuss privacy, compliance, and misuse risks, and release a fully simulated dataset with no raw transaction records or personally identifiable information, balancing scientific utility with responsible data stewardship.

  1. The controlled ablations comparing original and simplified graph constructions, along with the separation of user-level and account-level tasks, make the empirical questions transparent and easy to follow.
Critical weaknesses
  1. The dataset is generated using an off-the-shelf simulation framework (Darabi et al., 2025) with standard CTGAN feature kernels, and the modeling experiments consist entirely of existing GNN baselines with no new algorithmic contribution. The core empirical finding that heterogeneous GNNs outperform homogeneous GNNs on a heterogeneous graph is entirely expected and does not demonstrate unique scientific value beyond the dataset itself.

  2. The authors explicitly withhold the full data generation process, exact hyperparameters, computational environment details, and complete simulation code for ethical reasons. While the rationale is understandable, this severely undermines the paper’s central claim of improving “openness, reproducibility, and fair comparison.” A benchmark whose construction cannot be independently verified or replicated has limited long-term scientific value.

  3. The comparison with specialized AML models is flawed: LineMVGNN fails due to out-of-memory errors on a single 32 GB GPU, suggesting inadequate experimental engineering rather than meaningful dataset difficulty; MEGA-GNN and DGA-GNN are designed for Bitcoin/fraud networks, not banking graphs. Concluding that “current leading AML models bring only limited gains” from this comparison is misleading and methodologically weak.

  4. The AO/EO overlap analysis and the structural characterizations are purely observational correlations without causal or predictive validation. The paper does not show, for example, that entity-level overlap features actually improve a deployed detection pipeline, nor does it demonstrate cross-dataset transfer or generalization of methods trained on PERISCOPE to other AML benchmarks.

Questions for authors

see weakness

Reviewer_pAWaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents PERISCOPE, an AML benchmark built by simulating a single real bank’s data. The released graph has four node types (in-bank users, in-bank accounts, counterparty entities and counterparty accounts), two relations (ownership and timestamped transactions), and labels at both account and user level that come from real investigations, widened by a broader criterion for alert-worthy accounts. The paper describes how the graph was built and checks its fidelity against the original data, after comparing three feature kernels and picking CTGAN. It then runs two studies: a comparison of the heterogeneous graph against a collapsed variant across nine models, and a benchmark of three AML-specific homogeneous models against generic baselines. A separate analysis looks at overlaps among suspicious multi-account users.

Critical strengths

S1. The dataset fills a real gap — existing AML benchmarks collapse banking into a single-relation transfer graph, and PERISCOPE keeps the full bank schema.

S2. The reporting is candid about negative results: LineMVGNN running out of memory on the dense dataset, and the EO-only lift of 0.90 reported even though it works against the paper's own narrative.

Critical weaknesses

The headline claim contradicts Table 4. The abstract says the richer structure consistently improves performance over simplified transfer-graph constructions, but by AUPRC in Table 4 four of the nine user-level models do better in the simplified setting, and two account-level models improve as well. Section 5.1 only claims the defensible version — best-on-original against best-on-simplified. So the abstract and introduction claim more than the experiments show; the benefit is model-dependent, not consistent.

The comparison does not isolate the paper's central signal. The paper argues that cross-institution information is hard to obtain, but the collapsed variant in Table 4 still keeps the external-entity information — at the user level, transactions are mapped directly to outer entities — so no experiment removes it and tests what it contributes, and Appendix B reports an EO-only lift of 0.90. The baselines also leave out the models behind the benchmarks it compares against, including BWGNN, the model from which T-Finance is taken, and the graph fraud detection line this setting builds on. Section 5.2 dismisses heterogeneous AML models on grounds specific to the Bitcoin network, which does not cover the work the authors cite. And the abstract says AML-specific models bring only limited gains over generic graph baselines, while Section 5.2 reports that they consistently outperform them — the two statements refer to the same models and do not agree.

The main contribution is on the dataset and evaluation side rather than in the method. Judged as a dataset contribution, it would be acceptable once the claim is properly restated.

Questions for authors

Q1. The abstract says retaining the richer structure consistently improves AML performance. In Table 4, four of the nine user-level models do better in the simplified setting by AUPRC (MLP 47.50 to 55.43, GIN 59.39 to 62.99, RGCN 53.61 to 58.53, RGAT 54.43 to 59.96), and two account-level models improve as well, while Section 5.1 only reports best-on-original against best-on-simplified, 66.93 against 63.55. Was "consistently" meant in that best-against-best sense alone? If the benefit is model-dependent, what should the paper report so that a practitioner can tell when the richer construction pays off?

Q2. Section 3.1 says cross-institution associations are hard to obtain and often only partially observable, yet the collapsed variant in Table 4 still carries the external-entity information, and Appendix B reports an EO-only lift of 0.90. Is there any setting in which external-entity ownership is removed or partially withheld? And how would the main conclusions change if it were?

Q3. Are the AO/EO statistics in Figures 3 and 4 and Appendix B computed on the simulated graph or on the raw data? If on the simulated graph, the simulation pipeline — the structure generator and the post-hoc label refinement — may itself induce the overlap pattern; do the corresponding statistics on the raw data, even as aggregate ratios or percentiles, rule this out?

Q4. Section 5.2 excludes heterogeneous AML models on the grounds that they are designed for the Bitcoin network rather than real bank data, but this does not cover the work the authors cite, and BWGNN from Tang et al. 2022 sits in the reference list while its models are not evaluated, as is the graph fraud detection line this setting builds on (CARE-GNN, DiG-In-GNN and others). What is the basis for the general conclusion that AML-specific models bring only limited gains?

Reviewer_FtVcOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces PERISCOPE, a simulation-based anti-money laundering (AML) benchmark grounded in real banking data that preserves customer–account ownership relations, external-entity associations, and transaction timestamps. Through user- and account-level evaluations of existing graph models, the authors report that retaining this richer structure improves detection performance compared with simplified transfer graphs.

Critical strengths
  1. PERISCOPE preserves customer–account ownership relations and external-entity associations that are often omitted from simplified AML transaction graphs. This provides a potentially useful resource for studying how these relationships contribute to AML detection.

  2. The paper evaluates existing graph models at both the user and account levels, providing initial reference results for future research using the dataset.

Critical weaknesses
  1. The experiments do not directly establish the detection benefit of external-entity associations. These associations are a central contribution of PERISCOPE, but Table 4 changes several graph components simultaneously, along with model depth and sampling strategies. It therefore does not isolate their contribution. Furthermore, Appendix B reports an EO-only lift of 0.90, meaning that the additional account pairs identified through entity-level overlap are not enriched for suspicious users in the sampled evaluation set. This does not imply that these associations are useless, but their value when combined with other information still needs to be demonstrated. A controlled comparison that removes only external entity–account ownership relations would provide more direct evidence.

  2. The training, validation, and test protocol for node classification is unclear. The paper does not clearly specify which node labels are used for training and evaluation, whether the same labeled nodes appear across splits, or how labels align with prediction times. Without these details, it is difficult to determine exactly what the reported performance measures and whether information leakage is adequately prevented.

  3. Synthetic label generation and subsequent adjustments are insufficiently explained. Section 3.1 states that the original labels come from real investigations, while Appendix A.2 mentions further refinement of generated features and labels. However, the paper does not explain how newly generated accounts receive labels or what rules are used to adjust them afterward. Although the statistical comparisons and experiments on real banking data provide useful validation, they do not resolve this ambiguity. Since labels define the detection task, these details are necessary to assess whether models learn realistic risk patterns or patterns introduced by the generation and adjustment procedures.

  4. The analysis provides limited insight beyond descriptive observations. The main findings—that suspicious accounts tend to have higher transaction activity, that entity-level overlap covers more account pairs than account-level overlap, and that selected heterogeneous models benefit from richer graph structure—are largely expected. In particular, broader EO coverage follows from AO being a subset of EO by definition. The analysis does not sufficiently establish whether external-entity associations remain informative after controlling for transaction activity and account connectivity, or identify when and why these associations improve detection. Consequently, the claimed domain-informed insights offer limited guidance beyond the general observation that relational information can be useful for AML.

Questions for authors
  1. External-entity associations: Could you report a controlled ablation that removes only external entity–account ownership relations while keeping other graph components, features, and model settings unchanged?

  2. Synthetic labels: How are labels assigned to newly generated accounts? Appendix A.2 mentions further refinement of generated features and labels; what adjustments are made, and what information or rules guide them?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.