A Privacy-Preserving Federated Framework with Hybrid Quantum-Enhanced Learning for Financial Fraud Detection

Abhishek Sawaika, Swetang Krishna, Tushar Tomar, Durga Pritam Suggisetti, Aditi Lal, Tanmaya Shrivastav, Nouhaila Innan, Muhammad Shafique

Introduction

Financial fraud detection is a critical yet challenging task, characterized by a constant "cat-and-mouse" game between institutions and sophisticated actors. Identifying fraudulent patterns requires analyzing massive, high-dimensional datasets to uncover subtle, cross-transactional correlations. However, strict privacy regulations, such as the General Data Protection Regulation (GDPR), prevent financial institutions from pooling sensitive customer data into centralized repositories.

While Federated Learning (FL) provides a decentralized training alternative, it is often hindered by the inability of classical models to capture complex sequential patterns and the vulnerability of the FL process to poisoning and inference attacks. This paper proposes a novel framework that integrates Quantum-enhanced Long Short-Term Memory (QLSTM) networks with a specialized privacy mechanism called FedRansel to address these limitations.

The Hybrid QLSTM Architecture

The core of the proposed system is the QLSTM, which replaces the classical neural network gates (forget, input, and output) with Variational Quantum Circuits (VQCs). This hybrid approach leverages quantum mechanics to enhance feature representation:

  • Encoding: Classical transaction data is mapped into a high-dimensional Hilbert space using angle encoding (RX rotations).
  • Entanglement: The architecture utilizes CNOT gates to create quantum entanglement, allowing the model to capture complex relationships between features that are often missed by classical bits.
  • Processing: Parameterized "Rot" gates are optimized during training, utilizing quantum parallelism to explore the solution space more efficiently for fraud detection tasks.

Research indicates that these quantum layers are particularly effective at identifying "cross-transactional patterns"—the sequential links between multiple transactions that signal fraudulent behavior.

FedRansel: Securing the Federated Process

To protect the integrity of the model and the privacy of the participants, the authors developed FedRansel (Federated Random Selection). Unlike standard FL, where clients transmit their entire model updates, FedRansel requires clients to share only a randomly sampled subset of parameters.

The global server only updates parameters that meet a specific sharing threshold across nodes. This "pseudo-centralized" strategy ensures that the server never possesses a complete view of any single client's model, making it significantly harder for attackers to perform model inversion or membership inference attacks. Notably, FedRansel avoids the significant accuracy degradation often associated with traditional Differential Privacy (DP) techniques.

Experimental Results and Discussion

The framework was evaluated using the PaySim synthetic financial dataset and a real-world Bank Fraud Detection dataset. To accommodate the constraints of current quantum simulators (PennyLane), Principal Component Analysis (PCA) was used to reduce data dimensionality while preserving essential variance.

Performance Gains

  • Accuracy and Recall: The QLSTM outperformed classical LSTMs by approximately 5% in accuracy and 10% in recall on the synthetic dataset.
  • Attack Resilience: Under poisoning and inference attacks, FedRansel reduced model degradation by 4–8% compared to standard DP methods.
  • Privacy Utility: The framework showed an 8% improvement in resisting membership inference attacks with only a negligible (~1%) drop in overall predictive performance.

Optimal Complexity

The study found that "over-quantizing" does not necessarily lead to better results. For the real-world dataset, performance peaked at 5 qubits and 4 layers, suggesting that maintaining an optimal level of circuit complexity is vital to avoid overfitting and optimization hurdles like barren plateaus.

Limitations and Future Directions

While the results are promising, several challenges remain for real-world deployment:

  1. Scalability: Model performance tends to degrade as the number of participating nodes increases, likely due to data sparsity at the local level.
  2. Hardware Constraints: The current study relies on quantum simulators. Transitioning to Noisy Intermediate-Scale Quantum (NISQ) hardware will introduce decoherence and hardware noise.
  3. Training Stability: Like many quantum neural networks, the model is susceptible to vanishing gradients (barren plateaus) as circuit depth increases.

Future work will focus on testing the framework on physical quantum processors and exploring quantum-native communication channels to further secure federated nodes.

Summary updated

BibTeX
@misc{sawaika2025privacy,
  author = {Abhishek Sawaika and Swetang Krishna and Tushar Tomar and Durga Pritam Suggisetti and Aditi Lal and Tanmaya Shrivastav and Nouhaila Innan and Muhammad Shafique},
  title = {{A Privacy-Preserving Federated Framework with Hybrid Quantum-Enhanced Learning for Financial Fraud Detection}},
  year = {2025},
  eprint = {2507.22908},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.22908},
}
Figure