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

TL;DR

This work tackles privacy-preserving fraud detection in finance by fusing quantum-enhanced sequential learning with federated training and a novel FedRansel privacy mechanism. It introduces a quantum-enhanced LSTM (QLSTM) whose gates are implemented via variational quantum circuits and integrates this within a pseudo-centralized federated learning framework to bolster robustness against poisoning and inference attacks. Empirical results on synthetic and real banking datasets show a ~5% gain over classical LSTMs and ~4–6% improvements in privacy-related robustness over differential privacy baselines, validating the approach on quantum simulators. The framework demonstrates favorable trade-offs between privacy, security, and learning performance, with implications for scalable, privacy-aware quantum-enabled analytics in finance and other sensitive domains.

Abstract

Rapid growth of digital transactions has led to a surge in fraudulent activities, challenging traditional detection methods in the financial sector. To tackle this problem, we introduce a specialised federated learning framework that uniquely combines a quantum-enhanced Long Short-Term Memory (LSTM) model with advanced privacy preserving techniques. By integrating quantum layers into the LSTM architecture, our approach adeptly captures complex cross-transactional patters, resulting in an approximate 5% performance improvement across key evaluation metrics compared to conventional models. Central to our framework is "FedRansel", a novel method designed to defend against poisoning and inference attacks, thereby reducing model degradation and inference accuracy by 4-8%, compared to standard differential privacy mechanisms. This pseudo-centralised setup with a Quantum LSTM model, enhances fraud detection accuracy and reinforces the security and confidentiality of sensitive financial data.

Figures (9)

• Figure 1: Overview of FL, each participating organization trains a local model on private data, while a central federated server aggregates these local updates to construct a collaborative global model without data sharing.
• Figure 2: Common attacks and their target realization in FL. The diagram represents a traditional FL setup, including various types of attacks that occur during both the training and prediction phases. It indicates the exact location/step where the attack happens in the overall system. Our focus is on the ones indicated by the arrows during the training phase. $\Delta W's$ here represent the model updates by individual participants.
• Figure 3: Training workflow of our framework demonstrating local quantum-enhanced computations, federated aggregation, and secure update exchanges.
• Figure 4: QLSTM architecture overview. A three-sequence QLSTM model where each QLSTM cell processes an input at time step $\mathbf{X}_t$, computes a corresponding hidden state $\mathbf{h}_t$, and generates the output $\mathbf{Y}_{t+1}$. This diagram illustrates the sequential stacking of QLSTM cells, facilitating both temporal and federated model integration for enhanced dynamic learning capabilities.
• Figure 5: The quantum circuit, composed of a sequence of RX operations that implements angle encoding, while the parameterized Rot gate and CNOT operations create a fully entangled variational circuit.