Scaling Trust in Quantum Federated Learning: A Multi-Protocol Privacy Design

TL;DR

Addresses privacy challenges in quantum federated learning by proposing a multiprotocol framework combining SVD, QKD, AQGD, DP, PCA-DP, data condensation, and pruning. Provides theoretical analysis and comprehensive experiments demonstrating privacy guarantees without prohibitive loss in training efficiency. Shows that multi-layer privacy mechanisms effectively safeguard data and model confidentiality while maintaining competitive performance and reduced communication overhead. Highlights potential for practical deployment of privacy-preserving QFL on near-term quantum devices.

Abstract

Quantum Federated Learning (QFL) promises to revolutionize distributed machine learning by combining the computational power of quantum devices with collaborative model training. Yet, privacy of both data and models remains a critical challenge. In this work, we propose a privacy-preserving QFL framework where a network of $n$ quantum devices trains local models and transmits them to a central server under a multi-layered privacy protocol. Our design leverages Singular Value Decomposition (SVD), Quantum Key Distribution (QKD), and Analytic Quantum Gradient Descent (AQGD) to secure data preparation, model sharing, and training stages. Through theoretical analysis and experiments on contemporary quantum platforms and datasets, we demonstrate that the framework robustly safeguards data and model confidentiality while maintaining training efficiency.

Figures (9)

• Figure 1: Overview of the proposed privacy-preserving QFL framework. A network of $n$ quantum devices trains local models and transmits them to the server after applying privacy mechanisms across data preparation, model sharing, and training stages. The design integrates Singular Value Decomposition (SVD), Quantum Key Distribution (QKD), and Analytic Quantum Gradient Descent (AQGD) to enable a robust multiprotocol privacy layer.
• Figure 2: Privacy-Preserving QFL Framework: Various protocols are implemented to provide further privacy to the QFL framework that include privacy during training, data privacy and communication privacy; ① Data condensation, ② PCA DP, ③ AQGD DP, ④ Noise, ⑤ Pruning, ⑥ DP, ⑦ SVD + QKD
• Figure 3: Implementation of SVD based privacy secure QFL framework.
• Figure 4: Model pruning approach for QFL privacy
• Figure 5: Global Model Adaptation (G+), Prediction (Pred.), Average Devices Performance (Avg.): PCA vs PCA_DP

Theorems & Definitions (1)

• Definition 1: ($\varepsilon, \delta$)-Differential Privacy