PaperlayerView exact revision on arXiv

Source-linked AI summary

Collaborative Synthetic Data for Privacy-Preserving Financial Fraud Detection Across Organizational Silos

Simeon Allmendinger, Domenique Zipperling, Burhanettin Bahadir Kibar, Niklas Kühl

arXiv:2609.33754v2cs.LGcs.AI

TL;DR

Fragmented organizational data and privacy constraints complicate AI analytics, while fraud detection is further challenged by scarce minority-class examples. The paper adapts collaborative diffusion-based synthetic data generation to tabular fraud detection. Its approach is not the most locally faithful, but improves detection more consistently across most datasets.

  • Problem

    Fragmented data, privacy constraints, and scarce informative minority-class observations limit inter-organizational analytics for rare-event tasks such as fraud detection.

    Source: Introduction · pp. 1–2

  • Method

    The paper adapts collaborative diffusion to tabular fraud data, splitting denoising work between a server and organizations that retain raw records locally.

    Source: Training a Collaborative Diffusion Model · p. 7

  • Results

    Collaborative diffusion does not produce the most locally faithful fraud samples, yet improves fraud detection more consistently across most datasets than local interpolation-based methods.

    Source: Discussion · p. 12

  • Takeaways & Limitations

    Synthetic data quality for organizational analytics should be assessed by downstream utility and governance fit as well as local distributional fidelity.

    Source: Discussion · p. 13

    The study does not report empirical wall-clock training times, memory footprints, or communication volumes, and its findings are specific to tabular fraud data.

    Source: Limitations · p. 14

Abstract

from arXiv · show

Organizations seek analytical value from AI, yet relevant data are often fragmented across organizations and constrained by privacy. This is acute in financial fraud detection, where rare fraud cases and imbalanced local datasets limit decision-relevant analytics. Federated learning enables collaboration without direct data sharing but does not resolve minority-class scarcity. Synthetic data generation can help, yet lightweight methods are interpolation-bound, while generative models require substantial data and computation. Existing collaborative generative approaches often rely on federated learning, imposing considerable organization-side training burdens. In this paper, we examine CollaFuse as a collaborative diffusion-based alternative for fraud detection and evaluate it across five fraud datasets. Compared with classical oversampling, local generative baselines, and centralized diffusion benchmarks, CollaFuse does not achieve the highest local fidelity but improves downstream fraud detection more consistently across most datasets. These findings suggest that synthetic data create analytical value less through local realism than through transferable cross-organizational structure.

Introduction

Data fragmentation and privacy constraints hinder organizations’ ability to mobilize data for AI, while rare-event domains such as fraud detection face a scarcity of informative minority-class observations. The section motivates collaborative diffusion-based synthetic data generation as a way to create analytical value from distributed, imbalanced data without sharing raw records.

Introduction

  • Data fragmentation and limited central access hinder AI implementation, while sharing feasibility also depends on privacy, governance, trust, sovereignty, incentives, readiness, and coordination.Relevant data are described as critical organizational resources for effective AI development.Source: Introduction · p. 1
  • Federated learning supports inter-organizational analytics without raw-data sharing, but does not resolve scarcity of informative minority-class observations.This limitation matters in rare-event domains such as fraud detection and healthcare.Source: Introduction · p. 2
  • Synthetic data methods face a trade-off: interpolation-based oversampling struggles with complex distributions, while generative approaches require substantial data and computation.Federated generative methods also impose local computation and memory burdens because each organization trains the entire model.Source: Introduction · p. 2
  • Collaborative diffusion-based generation is adapted to tabular fraud detection to examine whether it can create analytical value from distributed, imbalanced, privacy-constrained data.The study evaluates fidelity to fraud-related distributions and usefulness for downstream detection across classifier types.Source: Introduction · p. 2
  • The paper argues that collaborative synthetic data can create value without closely reproducing local distributions, by transferring task-relevant cross-organizational patterns without pooling raw data.It frames this as a methodological adaptation and a conceptual account of synthetic data as a decision-relevant analytical resource.Source: Introduction · p. 2

Background and Related Work

Fraud analytics must create value from rare, fragmented data without pooling raw records. Prior methods address pieces of this challenge, but whether collaborative synthetic data transfer task-relevant structure across organizations remains unestablished.

Background and Related Work

Fraud Detection Under Class Imbalance

Traditional Method to Create Synthetic Data for Minority Fraud Cases

Generative Methods for Synthetic Minority Fraud Data

Privacy-Preserving Collaborative Diffusion Learning

  • Timestep-split collaborative diffusion assigns earlier denoising to a server and final steps to organizations, addressing client-side burdens while retaining local control.At effective cut points, clients execute at most 20% of denoising steps.Source: Privacy-Preserving Collaborative… · pp. 5–6

Collaborative Diffusion for Privacy-Preserving Fraud Detection

CollaFuse adapts timestep-split collaborative diffusion to generate synthetic fraud samples for local augmentation while organizations retain raw data. Its evaluation spans five fraud datasets and compares fidelity and downstream detection utility across heterogeneous organization-level settings.

Experimental Setup

  • Across five fraud datasets, the pipeline evaluates collaborative training, local generation of synthetic fraud samples, and downstream detection against real-data-only and alternative augmentation baselines.The datasets and organization partitions are designed to capture variation in feature structure and organizational heterogeneity.Source: Collaborative Diffusion for… · p. 6, Stage 3: Downstream Fraud Detection… · p. 11

Training a Collaborative Diffusion Model

  • The server learns from later denoising steps using noisy intermediate representations, while organizations retain the more privacy-sensitive local part of the reverse process.Organizations transmit noised representations rather than raw transaction records; the cut point controls how much of the reverse process is handled locally versus by the server.Source: Training a Collaborative Diffusion Model · p. 7
  • The privacy argument assumes an honest-but-curious server and non-colluding organizations, but acknowledges residual leakage through noisy representations, memorization, and model probing.The paper does not claim formal differential privacy guarantees.Source: Training a Collaborative Diffusion Model · p. 7
  • A fraud-aware composite organization-side loss combines reconstruction, prior regularization, and a triplet term to guide generation toward fraud samples useful for minority-class augmentation.The triplet term encourages generated fraud samples to lie closer to real fraud observations than to non-fraud observations.Source: Training a Collaborative Diffusion Model · p. 7
  • After collaborative training, the server and each organization perform successive portions of the reverse process, returning generated fraud samples to their originating organization for local augmentation.Generation starts from Gaussian noise; synthetic samples are not pooled globally.Source: Training a Collaborative Diffusion Model · pp. 7–8

Experimental Setup

Results

Across most datasets, collaborative diffusion provides the strongest downstream fraud-detection utility, while local interpolation-based methods achieve closer alignment with client-level fraud distributions. The detection advantage is strongest on IEEE-CIS, Elliptic, BAF, and PaySim, and less pronounced on Credit Card Fraud.

Distributional Fidelity of Synthetic Fraud Samples

  • Local interpolation-based methods generally achieve the lowest MMD, indicating closer alignment with organization-level fraud distributions than collaborative diffusion.The pattern holds across datasets; fidelity among other generative baselines varies by dataset and organization composition.Source: Results · p. 12

Downstream Utility for Fraud Detection

  • Collaborative diffusion achieves the strongest downstream utility across most datasets, consistently outperforming the non-augmented baseline and exceeding SMOTE and ADASYN.The relative ranking remains largely stable across augmentation ratios and organization-level partitions, with results competitive but less pronounced on Credit Card Fraud.Source: Results · p. 12, Table 2 · p. 12
  • The strongest detection gains occur on IEEE-CIS, Elliptic, BAF, and PaySim, while Credit Card Fraud results remain competitive but less pronounced.Absolute F1 levels vary across datasets and organization-level partitions.Source: Results · p. 12

Discussion

Collaborative diffusion improves fraud detection more consistently across most datasets despite lower local fidelity, suggesting that transferable task-relevant structure can matter more than local realism. The discussion frames this value as dependent on collaboration architecture and complementary data, while noting scope and privacy limitations.

Discussion

  • Local interpolation methods often remain closer to observed fraud distributions, yet collaborative diffusion improves fraud detection more consistently across most datasets.The discussion concludes that locally faithful samples are not necessarily the most useful for downstream decisions.Source: Discussion · p. 12
  • Collaborative generation can transfer cross-organizational minority-class structure without pooling raw transaction records, but it provides no formal privacy guarantees.The authors describe it as an architecture that can partially relax data-access constraints, not eliminate the need for data or privacy analysis.Source: Discussion · p. 13
  • Fidelity and downstream utility can diverge, so synthetic-data quality should include task value, governance compatibility, and decision performance.The discussion argues that statistical resemblance alone captures only one dimension of quality.Source: Discussion · p. 13
  • Noisy shared representations support broader fraud-relevant structure, while final generation and downstream training remain local; benefits are weaker on Credit Card Fraud than on IEEE-CIS, Elliptic, BAF, and PaySim.The authors link usefulness to complementary minority-class structures across organizations, rather than decentralization alone.Source: Discussion · p. 13
  • The study contributes by extending evaluation beyond fidelity, framing synthetic data as cross-organizational knowledge transfer, and treating shared-versus-local denoising as a governance arrangement.The proposed evaluation considers fidelity, utility, governance fit, and privacy risk.Source: Discussion · p. 13
  • Organizations should evaluate synthetic-data methods across the full analytics pipeline and compare them with simpler baselines, rather than selecting by generative realism alone.The discussion names class weighting, random oversampling, SMOTE, and ADASYN as comparison baselines.Source: Discussion · p. 13
  • The timestep-split architecture limits local execution to the final denoising share and exchanges batch-sized noised representations, but synthetic data can still leak information.The authors recommend privacy testing, governance controls, and clear exchange rules; the efficiency argument is architectural rather than based on reported wall-clock times or resource measurements.Source: Discussion · p. 14, Limitations · p. 14

Limitations

  • The study does not fully benchmark other collaborative generative paradigms, reports no empirical runtime or memory measurements, uses few clients, and is limited to tabular fraud data.The authors caution that synthetic-data quality does not guarantee privacy and do not claim formal privacy guarantees.Source: Limitations · p. 14

Conclusion and Outlook

The conclusion presents collaborative diffusion as a way to mobilize rare, sensitive, and fragmented data for local decision-making. It reports improved local fraud detection through cross-organizational structure transfer without pooling raw records, while calling for evaluation beyond fidelity and further governance research.

Conclusion and Outlook

  • Collaborative diffusion is framed as an IT-enabled capability for turning rare, sensitive, fragmented data into decision-relevant analytical value.The conclusion grounds this framing in resource-based views and IT capability.Source: Conclusion and Outlook · p. 14
  • Collaborative synthetic data generation can improve local fraud detection by transferring task-relevant structures across organizational boundaries without pooling raw records.The conclusion links evaluation to downstream utility, governance compatibility, and inter-organizational value creation, not distributional fidelity alone.Source: Conclusion and Outlook · p. 14
  • Future research should examine privacy safeguards, usage rights, accountability, contribution valuation, and when structure transfer yields value rather than negative transfer.The authors also identify governance of synthetic artifacts across organizational boundaries as an open research direction.Source: Conclusion and Outlook · p. 14
Loading 2609.33754v2…