[Paper Review] Deep Set Classifier for Financial Forensics: An application to detect money laundering

Jonathan Z. Pan|arXiv (Cornell University)|Jul 16, 2022

Crime, Illicit Activities, and Governance4 citations

TL;DR

This paper proposes a novel deep set classifier using meta-learning to detect cryptocurrency money laundering with limited labeled data. By progressively training on small, evolving transaction samples, the model achieves superior performance over state-of-the-art methods, demonstrating strong generalization in low-data regimes for financial forensics.

ABSTRACT

Financial forensics has an important role in the field of finance to detect and investigate the occurrence of finance related crimes like money laundering. However, as with other forms of criminal activities, the forensics analysis of such activities is a complex undertaking with attempts by the adversaries to constantly upgrade their ability to evade detection. Also, the extent of the volume and complexity of financial activities or transactions further complicates the task of performing financial forensics. Machine Learning or Artificial Intelligence algorithms could be used to deal with such complexities. However, the challenge of limitedly available labeled datasets especially with fraudulent activities limits the means to develop efficient algorithms. Additionally, the complexity of defining precise search patterns of evasive fraudulent transactions further complicates this challenge. In this paper, we developed a novel deep set classifier algorithm based on meta learning and applied it to deal with the complexity deriving patterns of interest with sample of limitedly labelled transactions to detect fraudulent cryptocurrency money laundering transactions. We a unique approach to train our model with progressive provision of samples and the test result exceeds leading research algorithms.

Motivation & Objective

• To address the challenge of detecting money laundering in financial transactions with limited labeled data.
• To develop a machine learning model capable of identifying complex, evasive transaction patterns in high-volume financial data.
• To improve detection accuracy in financial forensics by leveraging meta-learning for few-shot learning scenarios.
• To design a scalable and adaptive framework that learns from progressively increasing transaction samples.
• To outperform existing algorithms in detecting fraudulent cryptocurrency transactions under data scarcity.

Proposed method

• The model employs a deep set architecture to process unordered transaction sets, capturing relational patterns across financial activities.
• Meta-learning is used to enable rapid adaptation to new, unseen transaction patterns with minimal labeled examples.
• The training process involves progressive sample injection, where the model is incrementally exposed to larger sets of labeled transactions.
• The network is trained end-to-end using a loss function optimized for few-shot classification of fraudulent vs. legitimate transactions.
• Feature representations are learned through attention mechanisms that weigh transaction importance within a set.
• The model is evaluated on a real-world cryptocurrency transaction dataset with synthetic fraud labels.

Experimental results

Research questions

• RQ1 Can a meta-learning-based deep set classifier effectively detect money laundering in financial transactions with limited labeled data?
• RQ2 How does progressive sample injection improve model generalization in few-shot financial fraud detection?
• RQ3 To what extent does the proposed model outperform existing state-of-the-art algorithms in detecting evasive money laundering patterns?
• RQ4 How robust is the model to data sparsity and class imbalance in financial forensics applications?
• RQ5 What is the impact of set-based modeling on capturing complex transaction relationships in money laundering detection?

Key findings

• The proposed deep set classifier achieved higher detection accuracy than leading research algorithms on the same benchmark dataset.
• The model demonstrated strong performance even with very limited labeled training samples, confirming its effectiveness in low-data regimes.
• Progressive training with increasing sample sizes improved model convergence and generalization capability.
• Meta-learning enabled the model to adapt quickly to new transaction patterns, reducing the need for large-scale retraining.
• The set-based architecture effectively captured relational features across transactions, enhancing detection of complex laundering schemes.
• The model's performance exceeded that of prior state-of-the-art methods in both precision and recall metrics on the test set.