Privacy Technologies for Financial Intelligence

TL;DR

This paper surveys privacy technologies for financial intelligence, arguing that cross‑organisation data sharing is essential to detect complex AML/CTF crimes. It maps three core computational challenges—Private Intersection Identification, Cross‑Institutional Pattern Detection, and Federated Model Learning—and explains how MPC, HE, hashing, FL, and DP can address them. Through real‑world case studies and industrial deployments, it demonstrates how privacy‑preserving information sharing (e.g., SPL handling, FinTracer, collaborative learning) can improve detection accuracy while maintaining privacy guarantees. The discussion highlights regulatory harmonisation, privacy‑utility trade‑offs, and practical guidance for implementing privacy technologies in AML/CTF workflows. Overall, the work underscores the potential and limits of privacy technologies to enhance financial crime deterrence while acknowledging the need for supportive policy reform and governance.

Abstract

Financial crimes like money laundering and terrorism financing can have significant impacts on society, including loss of trust in the integrity of the financial system, misuse and mismanagement of public funds, increase in societal problems like drug trafficking and illicit gambling, and loss of innocent lives due to terrorism activities. Effective detection of complex financial crimes remains a formidable challenge for regulators and financial institutions because the critical data needed to establish patterns and criminality are often dispersed across multiple organisations and cannot be linked due to privacy constraints around large-scale data matching. Recent advances in privacy and confidential computing technologies, which enable private and secure data analysis across organisations, offer a promising opportunity for regulators and the financial industry to come together to enhance their collaborative risk detection while maintaining privacy standards. This paper, through a survey of the financial intelligence ecosystem, seeks to identify opportunities for the utilisation of privacy technologies to improve the state-of-the-art in financial-crime detection.

Figures (6)

• Figure 1: The figure shows a simplified schematic view of a financial intelligence ecosystem.
• Figure 2: The entity-resolved data and risk indicators available to a typical financial intelligence agency.
• Figure 3: Privacy-enhanced version of Figure \ref{['fig:fc-current']}. LEAs = Law Enforcement Agencies; PAs = Partner Agencies; EOI = Entities of Interest
• Figure 4: An example federated learning (FL) architecture with $n$ data owners. In the setting, data owners only exchange the model gradient ($w$) with the aggregation server. In each training iteration, each data owner $i$ trains its own model with their local data and send the corresponding local model updates $w_i$ to the aggregation server. After aggregating all the received local model updates, the aggregations server returns the new global model updates ($W$) to each data owner.
• Figure 5: A simplified transaction network to demonstrate FinTracer. Accounts with cross-organisation transactions are linked by directed edges. There are four financial institutions A, B, C and D, each having a number of accounts. The account $a_1$ in A directly received NDIS payment.

Theorems & Definitions (1)

• Definition 1