Secure and Explainable Fraud Detection in Finance via Hierarchical Multi-source Dataset Distillation

TL;DR

This work tackles the need for explainable and privacy-preserving fraud detection in finance by introducing tree-region dataset distillation, which converts random-forest leaves into axis-aligned hyperrectangles and uniformly samples within them to create synthetic transactions. The resulting surrogate dataset is compact, auditable, and provides global rule summaries as well as per-sample rationales with calibrated uncertainty, enabling cross-institution collaboration without exposing original data. Empirical evaluation on the IEEE-CIS Fraud Detection dataset shows 85–93% data reduction with strong cross-cluster gains, while privacy assessments (including membership inference) indicate low memorization risk; filtering by disagreement further improves AUC to 0.687. The method offers a practical, regulator-friendly alternative to heavier generative or cryptographic approaches and points to extensions in differential privacy, mixed feature handling, and governance for multi-party analytics in regulated financial environments.

Abstract

We propose an explainable, privacy-preserving dataset distillation framework for collaborative financial fraud detection. A trained random forest is converted into transparent, axis-aligned rule regions (leaf hyperrectangles), and synthetic transactions are generated by uniformly sampling within each region. This produces a compact, auditable surrogate dataset that preserves local feature interactions without exposing sensitive original records. The rule regions also support explainability: aggregated rule statistics (for example, support and lift) describe global patterns, while assigning each case to its generating region gives concise human-readable rationales and calibrated uncertainty based on tree-vote disagreement. On the IEEE-CIS fraud dataset (590k transactions across three institution-like clusters), distilled datasets reduce data volume by 85% to 93% (often under 15% of the original) while maintaining competitive precision and micro-F1, with only a modest AUC drop. Sharing and augmenting with synthesized data across institutions improves cross-cluster precision, recall, and AUC. Real vs. synthesized structure remains highly similar (over 93% by nearest-neighbor cosine analysis). Membership-inference attacks perform at chance level (about 0.50) when distinguishing training from hold-out records, suggesting low memorization risk. Removing high-uncertainty synthetic points using disagreement scores further boosts AUC (up to 0.687) and improves calibration. Sensitivity tests show weak dependence on the distillation ratio (AUC about 0.641 to 0.645 from 6% to 60%). Overall, tree-region distillation enables trustworthy, deployable fraud analytics with interpretable global rules, per-case rationales with quantified uncertainty, and strong privacy properties suitable for multi-institution settings and regulatory audit.

Figures (7)

• Figure 1: Illustration of tree-based dataset distillation.
• Figure 2: Visualization of the dataset partitioned into three clusters using $k$-means.
• Figure 3: Disagreement score distribution of synthesized positive and negative samples.
• Figure 4: Grid search over positive and negative disagreement percentile thresholds. Cell values denote downstream AUC.
• Figure 5: Test-set AUC as a function of the training-data distillation ratio.