[Paper Review] Privacy-Preserving Credit Card Fraud Detection using Homomorphic Encryption

David Nugent|arXiv (Cornell University)|Nov 12, 2022

Cryptography and Data Security4 citations

TL;DR

This paper proposes a privacy-preserving credit card fraud detection system using homomorphic encryption (HE), enabling secure inference on encrypted transaction data without exposing sensitive information. It implements and compares HE-encrypted XGBoost and feedforward neural network models, finding that XGBoost achieves significantly faster inference (6ms vs. 296ms) while maintaining strong detection performance, demonstrating feasibility for real-world deployment with privacy guarantees.

ABSTRACT

Credit card fraud is a problem continuously faced by financial institutions and their customers, which is mitigated by fraud detection systems. However, these systems require the use of sensitive customer transaction data, which introduces both a lack of privacy for the customer and a data breach vulnerability to the card provider. This paper proposes a system for private fraud detection on encrypted transactions using homomorphic encryption. Two models, XGBoost and a feedforward classifier neural network, are trained as fraud detectors on plaintext data. They are then converted to models which use homomorphic encryption for private inference. Latency, storage, and detection results are discussed, along with use cases and feasibility of deployment. The XGBoost model has better performance, with an encrypted inference as low as 6ms, compared to 296ms for the neural network. However, the neural network implementation may still be preferred, as it is simpler to deploy securely. A codebase for the system is also provided, for simulation and further development.

Motivation & Objective

• To address the privacy and security risks in traditional credit card fraud detection systems that process sensitive customer transaction data in plaintext.
• To design and implement a system enabling private fraud detection using homomorphic encryption (HE) for secure inference on encrypted transactions.
• To evaluate the performance, latency, and feasibility of deploying HE-encrypted XGBoost and neural network models in a real-world fraud detection context.
• To provide a publicly available codebase for simulation and further development of privacy-preserving fraud detection systems.

Proposed method

• Trained XGBoost and feedforward neural network models on plaintext credit card transaction datasets (ULB and Vesta) using standard machine learning pipelines.
• Converted the trained models into homomorphic encryption-compatible versions using the CKKS and Paillier schemes, enabling computation on encrypted inputs.
• Employed TenSEAL, a PyTorch-compatible HE library, to implement and benchmark encrypted inference for both models.
• Used order-preserving encryption (OPE) for XGBoost tree comparisons and CKKS for neural network layers to support homomorphic operations.
• Simulated the system with a client-host architecture where the client encrypts transactions and the host performs inference without accessing plaintext data.
• Evaluated model performance using standard fraud detection metrics (AUC, F1, precision, recall) and benchmarked latency and storage overhead on encrypted data.

Experimental results

Research questions

• RQ1 Can homomorphic encryption be effectively used to enable private inference for credit card fraud detection without compromising model performance?
• RQ2 How do the latency and storage overheads of HE-encrypted XGBoost and neural network models compare in a realistic fraud detection setting?
• RQ3 What is the trade-off between detection accuracy and inference speed when using HE-encrypted XGBoost versus neural networks on real-world transaction data?
• RQ4 How does the system handle information leakage risks, particularly in HE-encrypted XGBoost due to OPE-based comparisons?
• RQ5 What are the practical deployment challenges and potential optimizations for scaling HE-based fraud detection in production environments?

Key findings

• The HE-encrypted XGBoost model achieved an inference latency as low as 6ms, significantly outperforming the HE-encrypted neural network, which required 296ms.
• On the ULB dataset, the XGBoost model achieved an F1 score of 0.87 and AUC of 0.92, demonstrating strong detection performance.
• The neural network model had higher storage overhead due to the need to store encrypted weights and intermediate activations, but offered a simpler deployment model.
• The XGBoost model showed better detection performance than the neural network on both the ULB and Vesta datasets, particularly in reducing false negatives.
• Despite higher latency, the neural network implementation may be preferred in practice due to its simpler and more secure deployment pipeline.
• The study identified potential information leakage in HE-XGBoost due to OPE-based comparisons, suggesting the need for secure training or model obfuscation.