Authors:

Daniel Anoruo

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

As machine learning continues to advance in fields like healthcare and chemical engineering, concerns regarding data privacy and data breaches have been a topic of discussion for centralized machine learning models. Many legal regulations have been set in place to protect sensitive information across many domains, making it difficult for data to be shared among different institutions. To train models while addressing concerns around data privacy, Federated Learning (FL) was introduced as a method of decentralizing the data and enabling training locally among clients. During training, the model updates instead of the models themselves are shared back to the central server before aggregating into a refined version. Despite this approach being safe for protecting sensitive information, not only is FL slow from the distributed training, it is still vulnerable to attacks like model inversion and eavesdropping. To tackle these limitations, our research integrates federated learning with quantum computing and fully homomorphic encryption (FHE) to enhance both performance and security. We developed QFedLib, a library that combines the quantum framework Pennylane (Bergholm, 2018) with the PyTorch-based FL library Flower and embeds custom FHE capabilities for secure aggregation (Beutel et al., 2020). We also benchmarked our system on medical imaging tasks, including chest X-ray classification using the NIH dataset. Results showed that our FL model outperformed a centralized baseline in average accuracy. When incorporating quantum layers, training speed increased, though with a slight drop in performance likely due to current simulation constraints (Dutta, 2024). We plan on releasing our library as an open-source toolkit to support future research in privacy. Keywords: Quantum; Federated Learning (FL); Data Privacy; Fully Homomorphic Encryption

Source:

Purdue University / 2025

Topics:

No topics listed

Co-authors:

Daniel Anoruo