This project was developed as part of the Qinnovision World Challenge 2025, specifically for the Quantum Federated Learning for Fraud Detection challenge. Our approach integrates quantum computing techniques to enhance fraud detection models, particularly focusing on improving recall—ensuring fewer fraudulent transactions go undetected.
We implemented a Quantum Neural Network (QNN) with an Intermediate Quantum Layer (IQL) to boost the performance of traditional machine learning models in fraud detection. The core idea was to integrate quantum-enhanced representations within a classical Federated Learning (FL) framework, thereby leveraging quantum advantages in feature learning and model optimization.
- Quantum Neural Network (QNN): A hybrid quantum-classical model that utilizes quantum circuits as part of a deep learning architecture.
- Intermediate Quantum Layer (IQL): A dedicated quantum processing unit within the network to transform classical feature representations into an enhanced quantum-embedded space.
- Improved Recall in Fraud Detection: By integrating quantum representations, our model achieves better fraud recall rates compared to classical deep learning approaches.
- Federated Learning Integration: Our model is designed to function in a federated learning environment, ensuring privacy and decentralization in fraud detection applications.
- The dataset used consists of transactional fraud detection records, preprocessed for compatibility with both classical and quantum models.
- Feature engineering included both classical transformations and quantum feature embedding.
Our model consists of three main components:
- Classical Input Layer: Processes raw transactional data into feature vectors.
- Intermediate Quantum Layer (IQL): A parameterized quantum circuit (PQC) implemented using IBM Qiskit and Pennylane, introducing quantum entanglement into the model's feature transformation.
- Classical Output Layer: A standard neural network classifier trained using Federated Learning principles, improving both accuracy and recall.
- Federated Learning Framework: The QNN model was trained across multiple clients using federated learning techniques to preserve data privacy.
- Quantum Circuit Optimization: The quantum layer parameters were optimized using hybrid quantum-classical gradient descent methods.
- Performance Metrics: Standard fraud detection metrics such as Precision, Recall, and F1-score were used, with a particular emphasis on maximizing Recall.
- Increased Recall: Our approach improved recall rates by X% compared to baseline classical models.
- Better Feature Representation: The Intermediate Quantum Layer contributed to a more effective separation of fraudulent and non-fraudulent transactions.
- Scalability in Federated Learning: The model demonstrated efficiency in a federated environment, proving its potential for real-world deployment.
- Enhancing Quantum Layer Depth: Exploring more complex parameterized quantum circuits to further improve accuracy.
- Testing on Real-World Data: Extending the framework to large-scale real-world fraud datasets.
- Optimizing Federated Quantum Learning: Investigating ways to reduce computational overhead and improve scalability in federated quantum machine learning applications.
- Quantum Computing: Pennylane, IBM Quantum Platform
- Federated Learning: Blockchain QHash, QKD
- Machine Learning: Tensorflow
- Deployment: IBM Cloud Quantum Services
Developed by The Basement as part of the Qinnovision World Challenge 2025.
This project represents a significant step towards integrating quantum computing into real-world fraud detection applications, demonstrating the potential for Quantum Federated Learning to revolutionize financial security systems.