PaperlayerView exact revision on arXiv

Source-linked AI summary

Graph-Transformer Fraud Detection with Self-Supervised Pretraining and Conformal Risk Control

Sergei, Komarov

arXiv:2609.14234v1cs.LGcs.AI

TL;DR

Coordinated fraud rings are difficult to detect when systems inspect transactions independently. GTFD fuses graph and sequence representations with conformal risk control, self-supervised pretraining, and adversarial augmentation; on its corporate ledger benchmark, it reports strong detection and ring-recovery results.

  • Problem

    Fraud schemes form interlinked account rings, but transaction-by-transaction detection discards this structure, leaving coordinated money laundering and invoice fraud difficult to detect.

    Source: §I · p. 1

  • Method

    GTFD fuses multi-head graph attention and a gated sequence transformer, with conformal risk control, link-mask pretraining, and adversarial augmentation.

    Source: §I · p. 1, §VI · p. 7

  • Results

    On the investigator-verified corporate ledger benchmark, GTFD attains 0.990 AUROC and 96.1% F1-score, cuts false positives by about 29%, and recovers 96.5% of rings at a 5% false-positive budget.

    Source: §VI · p. 7

  • Takeaways & Limitations

    GTFD emits coverage-guaranteed scores that risk teams and regulators can reason about directly, and transfer experiments indicate the graph branch carries the largest share of its value.

    Source: L. Discussion · p. 7

    The investigator-verified synthetic rings may not capture all real criminal topologies, and privacy constraints were not evaluated.

    Source: L. Discussion · p. 7

Abstract

from arXiv · show

Financial fraud in corporate transaction networks has grown more coordinated and harder to detect with rule-based engines and with classical learning models that treat each transaction in isolation. This paper presents GTFD, a graph-transformer fraud detector that fuses structural and temporal evidence from a corporation's payment graph. GTFD encodes the graph with a multi-head graph attention network, encodes ordered transaction sequences with a gated transformer, and combines both views through a cross-modal gating layer. A conformal risk-control head converts the fused representation into threshold-free anomaly scores with finite-sample coverage guarantees, and the network is trained with self-supervised link-mask pretraining plus adversarial augmentation so it remains stable under scarce labels and under adversarial perturbation. On a corporate transaction benchmark enriched with coordinated fraud rings, GTFD reaches an AUROC of 0.990, an F1-score of 96.1% (precision 96.3%, recall 95.9%), and an accuracy of 98.4%. It reduces the false-positive rate by about 29% relative to the strongest baseline while raising coordinated fraud-ring recall from 85.1% to 96.5%. Ablations attribute roughly 2.0 AUROC points to self-supervised pretraining and 1.9 AUROC points to the conformal head, and adversarial stress tests show GTFD retains 89.2% accuracy at perturbation magnitude 0.20 where the next-best model falls to 76.4%.

I. INTRODUCTION

Coordinated fraud appears as rings of interlinked accounts, a structure missed by systems that inspect transactions independently. GTFD combines graph and sequence modeling with conformal scoring, self-supervised learning, and adversarial training.

I. INTRODUCTION

  • Corporate fraud schemes often form interlinked account rings, so transaction-by-transaction systems discard structure and miss coordinated money laundering and invoice fraud.The introduction links these blind spots to substantial institutional losses.Source: §I · p. 1
  • Prior work addresses relational patterns and label scarcity with graph learning, while sequence attention and conformal risk control address temporal patterns and threshold-free scores,,,.The supplied discussion presents GTFD as combining these strands.Source: §I · p. 1
  • GTFD reads transaction networks in graph and sequential forms, using multi-head graph attention, a gated transformer, and a learned cross-modal gate.The fused representation is paired with conformal scoring, link-mask self-supervision, and adversarial augmentation.Source: §I · p. 1
  • The stated contributions include jointly modeling structural and temporal signatures, providing coverage-guaranteed scores, learning from unlabeled ledgers, and hardening against worst-case perturbations.The paper says the self-supervised objective addresses the shortage of confirmed fraud labels.Source: §I · p. 1
  • The evaluated method beats both classical baselines and a recent graph–CNN hybrid by at least 1.6 AUROC points, cuts false positives by about 29%, and recovers 96.5% of coordinated fraud rings.Source: §I · p. 1

II. RELATED WORK

Prior fraud-detection approaches span rule-based and classical models, graph learning, and temporal attention, each addressing different aspects of the problem. Related work also covers pretraining, calibration, transfer, and explainability.

II. RELATED WORK

  • Rule-based systems are transparent but can produce false positives and miss emerging schemes, while classical learners rely on handcrafted features and ignore relational context,,.The review describes deep graph and sequence models as later additions to this progression.Source: §II · p. 1
  • Graph neural networks recover relational fraud patterns [1],, and self-supervised graph pretraining can reduce dependence on scarce labeled examples.Evidence-subgraph methods also expose the graph structure behind a score.Source: §II · p. 2
  • Temporal graph networks track streaming ledgers with moving time windows, and contrastive or generative pretext tasks can improve downstream accuracy from pretrained representations [15], [16].Source: §II · p. 2
  • Post-hoc explainers such as GNNExplainer and SHAP attribute scores to edges and features, supporting model audits,.The review connects such attribution to justifying alerts to regulators.Source: §II · p. 2
  • Transformers model long-range transaction dependencies, while gated token mixing and conformal risk control have been used to sharpen separation and produce calibrated prediction sets and p-values [4],,,.The review notes that calibration matters when alerts must be defensible to regulators.Source: §II · p. 2
  • The review places GTFD among graph–transformer hybrids and identifies conformal scoring, link-mask self-supervision, adversarial hardening, and coordinated-ring evaluation as distinguishing features.Source: §II · p. 2

III. PROPOSED METHOD

The method frames corporate fraud detection as scoring transfers and identifying accounts in coordinated fraud, despite few confirmed labels. Its graph branch uses attention to model relational signatures and retain evidence subgraphs.

A. Problem formulation

  • The task is to score each transaction and flag accounts participating in coordinated fraud using both graph structure and unlabeled sequence history despite few confirmed labels.The ledger is represented as a directed multigraph whose edges carry transaction attributes.Source: A. Problem formulation · p. 2

B. Graph branch

  • A two-layer multi-head graph attention network encodes each node’s structural representation from its incident edges.Source: B. Graph branch · p. 2
  • The graph branch captures dense, low-diameter fraud-ring clusters and retains an evidence subgraph with top contributing neighbors for flagged nodes.Source: B. Graph branch · p. 2

C. Sequence branch

The sequence branch processes each account’s ordered transaction stream with a gated token-mixing transformer. Its temporal embedding responds to velocity changes, round trips, and layering patterns.

C. Sequence branch

  • Each account’s ordered transaction stream is tokenized by amount, time delta, direction, and counterparty embedding before gated token mixing,.Source: C. Sequence branch · p. 2
  • The gated mixing operation interpolates between raw tokens and self-attention outputs to avoid over-smoothing rare, high-information events.The resulting temporal embedding responds to velocity changes, round-trip payments, and layering patterns.Source: C. Sequence branch · pp. 2–3
  • The pipeline combines the temporal transformer view with a structural graph-attention view through a cross-modal gate before conformal scoring.Source: Figure 1 · p. 3

G. Adversarial hardening

GTFD trains against plausible input perturbations, combines structural and temporal representations with a learned gate, and balances classification with auxiliary training objectives.

G. Adversarial hardening

  • Signed-gradient perturbations of input features train GTFD to keep scores stable when transaction amounts, timestamps, or narrative tokenization are changed.The perturbations are intended to represent a worst-case attacker and reach the maximum plausible extent.Source: G. Adversarial hardening · p. 3

H. Training objective

  • The training objective balances focal classification loss with masked-link reconstruction and adversarial regularization.The classification term dominates on the validation split while the auxiliary terms shape the representation.Source: H. Training objective · p. 3, D. Cross-modal fusion · p. 3

D. Cross-modal fusion

  • A learned gate combines structural and temporal embeddings, allowing the model to weight graph evidence for unusual account structure and sequence evidence for unusual behavior.The gate uses sigmoid weights to mix projected graph and sequence representations.Source: D. Cross-modal fusion · p. 3

E. Self-supervised link-mask pretraining

GTFD uses masked-link pretraining to learn from unlabeled transaction structure, then applies conformal risk control to produce calibrated, threshold-free transaction scores.

E. Self-supervised link-mask pretraining

  • Masked-link pretraining hides random edges and trains the encoder to reconstruct connections from neighboring structure before fraud supervision.The learned representation is intended to capture typical ledger topology and support fine-tuning with few confirmed labels.Source: E. Self-supervised link-mask pretraining · p. 3

F. Conformal risk control

  • A held-out calibration split yields a threshold-free conformal p-value for each transaction, supporting alerts with a finite-sample false-coverage guarantee.This avoids hand-tuning a decision cutoff.Source: F. Conformal risk control · p. 3

A. Datasets

The evaluation uses a chronologically split corporate transaction benchmark and compares GTFD with classical, graph, sequence, and hybrid baselines. GTFD leads every reported detection metric on the corporate test set.

A. Datasets

  • The primary benchmark contains 1.1M transfers among 240K accounts and 42 investigator-verified synthetic fraud rings; chronological 60/20/20 splits use past data to predict future fraud.PaySim and Elliptic provide additional benchmarks across mobile money and Bitcoin transactions.Source: A. Datasets · p. 3

B. Baselines

  • Baselines span classical, graph, sequence, and hybrid models, with classical models additionally evaluated using SMOTE and the better of their two configurations reported.Each model is repeated over three seeds, and means are reported.Source: B. Baselines · pp. 3–4

A. Comparison with baselines

  • GTFD leads every reported metric on the corporate test set, with 0.990 AUROC, 0.947 AUPRC, and 96.1% F1-score.Its AUROC exceeds the strongest baseline by 1.6 points and its AUPRC by 2.9 points; Table III presents the primary comparison.Source: A. Comparison with baselines · p. 4, Table III

B. Coordinated fraud-ring recovery

GTFD recovers coordinated fraud rings at low false-positive rates and dominates the baselines in ROC analysis. Ablations show contributions from both graph and sequence branches, pretraining, and conformal scoring.

B. Coordinated fraud-ring recovery

D. Ablation study

  • Removing graph or sequence branches costs 1.9 or 1.3 AUROC points; pretraining adds 1.1, conformal scoring adds 1.0 and lowers ECE from 4.8% to 1.7%, while adversarial training costs 0.8 clean-accuracy points.The graph and sequence views are complementary; adversarial training trades a small amount of clean accuracy for robustness.Source: D. Ablation study · pp. 4–5

C. ROC analysis

  • GTFD dominates baselines across the false-positive range, with 4.1% fraud misses and a 0.43% legitimate-account false-alarm rate.Figure 3 shows the ROC curves on the held-out test set; Figure 4 presents the normalized confusion matrix.Source: C. ROC analysis · p. 4, Figure 3 · p. 5, Figure 4 · p. 5

E. Adversarial robustness

GTFD retains accuracy under adversarial perturbation, maintains precision at high recall, and transfers across asset classes without retraining. Its explanations recover many investigator-annotated counterparties, while replacing its gated transformer with a temporal graph network lowers AUROC.

G. Cross-domain transfer

  • Without retraining across asset classes, GTFD reaches 0.921 AUROC on PaySim and 0.934 on Elliptic, retaining the most among the evaluated models.All models degrade when transferred from the corporate ledger, and the structural branch drives most of GTFD's transfer.Source: G. Cross-domain transfer · p. 5

H. Explainability and temporal encoding

  • Top explanation edges recover investigator-annotated counterparties at 87% fidelity, and replacing the gated transformer with a temporal graph network lowers AUROC by 1.1 points.The explanation evaluation uses a held-out set of 120 flagged accounts.Source: H. Explainability and temporal encoding · pp. 5–6

E. Adversarial robustness

  • At ε = 0.20, GTFD retains 89.2% accuracy versus 76.4% for the graph–CNN hybrid and 69.8% for a vanilla transformer under FGSM-style perturbations.Figure 6 reports accuracy as perturbation magnitude increases.Source: E. Adversarial robustness · p. 5, Figure 6 · p. 6

F. Precision–recall and calibration

I. Computational cost

GTFD's parallel branches and pretraining add deployment cost, though its throughput exceeds typical batch alerting needs and its overhead is confined to offline scoring. A recovered-ring case illustrates how an alert can expand to a ring, and hyperparameter perturbations produce small AUROC changes.

I. Computational cost

  • At 3.8 ms per transaction, GTFD sustains roughly 264K screenings per second on a single GPU, though it is not the cheapest model.The added latency accompanies its parallel branches and pretraining; the overhead is confined to offline scoring rather than real-time authorization.Source: §I · p. 6

J. Case study: a recovered invoice-fraud ring

  • In the case study, GTFD flags a hub in a seven-shell-account laundering ring and expands the alert to the ring while leaving two feeder accounts unflagged.Each shell account pockets 2–3% of a transfer before returning the remainder; the individual transfers appear plausible in isolation.Source: J. Case study: a recovered… · p. 6

K. Sensitivity to hyperparameters

  • Halving or doubling attention heads, transformer depth, or auxiliary-loss weights changes AUROC by at most 0.6 points.The authors report that the results are not an artifact of over-tuned hyperparameters.Source: K. Sensitivity to hyperparameters · p. 6

L. Discussion

GTFD combines structural and temporal fraud signals, with discussion findings highlighting complementary views, label-efficient pretraining, and coverage-guaranteed scoring. The conclusion reports strong corporate-ledger results and describes how scores and evidence subgraphs support investigation workflows.

L. Discussion

  • The graph and sequence branches are complementary, and the learned gate weights them differently for each account.Source: L. Discussion · p. 7
  • Link-mask self-supervision is described as the most cost-effective addition, transferring well to fine-grained fraud detection with few labels.Source: L. Discussion · p. 7
  • Conformal scores replace probabilities with implicit thresholds with coverage-guaranteed scores that risk teams and regulators can reason about directly.Source: L. Discussion · p. 7
  • The graph branch carries the largest share of cross-domain value, as ring and layering topology generalizes across asset classes despite differences in payment instruments.Source: L. Discussion · p. 7
  • Investigator-verified synthetic rings may not capture all real criminal topologies; privacy constraints were not evaluated.Source: L. Discussion · p. 7

VI. CONCLUSION

  • GTFD attains 0.990 AUROC, 96.1% F1-score, and 98.4% accuracy on a corporate ledger with investigator-verified fraud rings.The conclusion also reports about 29% fewer false positives and 96.5% ring recovery at a 5% false-positive budget.Source: §VI · p. 7
  • About 29% fewer false positives and 96.5% ring recovery at a 5% false-positive budget accompany transfer to mobile-money and cryptocurrency domains without retraining.Source: §VI · p. 7
  • The conformal score states a false-coverage rate, while evidence subgraphs name counterparties and ring-level signals let investigators expand an account into a laundering chain.Source: §VI · p. 7
Loading 2609.14234v1…