Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Fed-RD: Privacy-Preserving Federated Learning for Financial Crime Detection

Md. Saikat Islam Khan, Aparna Gupta, Oshani Seneviratne, Stacy Patterson

TL;DR

Fed-RD tackles end-to-end privacy in federated learning for relational financial data partitioned both vertically and horizontally.It introduces two fusion approaches: Approach 1 uses Local Rényi differential privacy with Gaussian noise on embeddings, while Approach 2 uses PBM and MPC to compute a private embedding sum; both yield formal $(\alpha, \epsilon)$-RDP guarantees over $Q$ iterations.The method models data with three components—a transaction model with parameters $\boldsymbol{\theta}_T$, an account model with parameters $\boldsymbol{\theta}_B$, and a fusion model with parameters $\boldsymbol{\theta}_F$—and optimizes the loss $\mathcal{L}(\boldsymbol{\\Theta}; \mathcal{X}_T, \mathcal{X}_B, \mathbf{y})$.Experiments on SWIFT-like and AMLSim synthetic datasets show that privacy-preserving Fed-RD attains high predictive performance while maintaining formal privacy guarantees and outperforming a baseline XGBoost model on transaction features.

Abstract

We introduce Federated Learning for Relational Data (Fed-RD), a novel privacy-preserving federated learning algorithm specifically developed for financial transaction datasets partitioned vertically and horizontally across parties. Fed-RD strategically employs differential privacy and secure multiparty computation to guarantee the privacy of training data. We provide theoretical analysis of the end-to-end privacy of the training algorithm and present experimental results on realistic synthetic datasets. Our results demonstrate that Fed-RD achieves high model accuracy with minimal degradation as privacy increases, while consistently surpassing benchmark results.

Fed-RD: Privacy-Preserving Federated Learning for Financial Crime Detection

TL;DR

Fed-RD tackles end-to-end privacy in federated learning for relational financial data partitioned both vertically and horizontally.It introduces two fusion approaches: Approach 1 uses Local Rényi differential privacy with Gaussian noise on embeddings, while Approach 2 uses PBM and MPC to compute a private embedding sum; both yield formal $(\alpha, \epsilon)$-RDP guarantees over $Q$ iterations.The method models data with three components—a transaction model with parameters $\boldsymbol{\theta}_T$, an account model with parameters $\boldsymbol{\theta}_B$, and a fusion model with parameters $\boldsymbol{\theta}_F$—and optimizes the loss $\mathcal{L}(\boldsymbol{\\Theta}; \mathcal{X}_T, \mathcal{X}_B, \mathbf{y})$.Experiments on SWIFT-like and AMLSim synthetic datasets show that privacy-preserving Fed-RD attains high predictive performance while maintaining formal privacy guarantees and outperforming a baseline XGBoost model on transaction features.

Abstract

We introduce Federated Learning for Relational Data (Fed-RD), a novel privacy-preserving federated learning algorithm specifically developed for financial transaction datasets partitioned vertically and horizontally across parties. Fed-RD strategically employs differential privacy and secure multiparty computation to guarantee the privacy of training data. We provide theoretical analysis of the end-to-end privacy of the training algorithm and present experimental results on realistic synthetic datasets. Our results demonstrate that Fed-RD achieves high model accuracy with minimal degradation as privacy increases, while consistently surpassing benchmark results.
Paper Structure (29 sections, 2 theorems, 12 equations, 4 figures, 3 tables, 2 algorithms)

This paper contains 29 sections, 2 theorems, 12 equations, 4 figures, 3 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Background
  3. Differential Privacy
  4. Gaussian Mechanism
  5. Poisson Binomial Mechanism (PBM)
  6. Multiparty Computation (MPC)
  7. Problem Formulation
  8. System Model
  9. Training Problem
  10. Threat Model and Privacy Objectives
  11. Proposed Algorithm
  12. Privacy
  13. Information Sharing
  14. Privacy Analysis of Approach 1
  15. Privacy Analysis of Approach 2
  16. ...and 14 more sections

Key Result

Theorem 5.1

Let $\beta' \in [0, \frac{1}{4}]$, $b' \in \mathbb{N}$, and $\alpha \leq 2$. Algorithm vflmpc, after $Q$ iterations, provides $(\alpha, \epsilon_T(\alpha))$-RDP for transactions features with and provides $(\alpha, \epsilon_B(\alpha))$-RDP for bank account features with where $P$ is the embedding size, $B$ is the transaction batch size, $M_T$ is the maximum number of transactions in which a sing

Figures (4)

  • Figure 1: Illustration of data and model partitioning among the transaction party and banks.
  • Figure 2: Testing AUPRC on the SWIFT dataset for various values of $\beta$, with $b = 64$ and $b' = 1024$.
  • Figure 3: Testing AUPRC on the AMLSim dataset for various values of $\beta$, with $b = 64$ and $b' = 1024$.
  • Figure 4: Comparison between Adam and SGD optimizer.

Theorems & Definitions (5)

  • Definition 2.1: Rényi divergence
  • Definition 2.2: Rényi Differential Privacy (RDP)
  • Definition 2.3: Local Rényi Differential Privacy (Local RDP)
  • Theorem 5.1
  • Theorem 5.2