FedQNN: Federated Learning using Quantum Neural Networks

TL;DR

This work thoroughly investigates QFL, underscoring its capability to secure data handling in a distributed environment and facilitate cooperative learning without direct data sharing and presents a novel framework to propel the field of QML into a new era of secure and collaborative innovation.

Abstract

In this study, we explore the innovative domain of Quantum Federated Learning (QFL) as a framework for training Quantum Machine Learning (QML) models via distributed networks. Conventional machine learning models frequently grapple with issues about data privacy and the exposure of sensitive information. Our proposed Federated Quantum Neural Network (FedQNN) framework emerges as a cutting-edge solution, integrating the singular characteristics of QML with the principles of classical federated learning. This work thoroughly investigates QFL, underscoring its capability to secure data handling in a distributed environment and facilitate cooperative learning without direct data sharing. Our research corroborates the concept through experiments across varied datasets, including genomics and healthcare, thereby validating the versatility and efficacy of our FedQNN framework. The results consistently exceed 86% accuracy across three distinct datasets, proving its suitability for conducting various QML tasks. Our research not only identifies the limitations of classical paradigms but also presents a novel framework to propel the field of QML into a new era of secure and collaborative innovation.

Figures (8)

• Figure 1: The general architecture of QML models starts with classical data $\mathbf{X}$, which undergoes data preprocessing and feature engineering. This data is converted into quantum-compatible states ${|\psi_j\rangle}$ via a mapping function $\phi$. For computation, these states are processed by a quantum circuit using quantum gates $f(\theta_n, X, R_{\ldots})$, including rotation and entangling gates. The model includes an iterative update loop where parameters are refined based on the lost function $J(\psi, y)$ until nearly converging ($J(\psi, y) \approx 0$). Finally, the quantum circuit's output is measured to produce the final model output $g$ for various predictive tasks.
• Figure 2: Network architecture for a typical FedML setup. Clients $\mathcal{C}_{i}$, for $1\leq i\leq k$, on local datasets $\mathcal{D}_{1},\mathcal{D}_{2},\ldots,\mathcal{D}_{k}$ with models $\mathcal{M}_{1},\mathcal{M}_{2},\ldots,\mathcal{M}_{k}$. The corresponding model parameters $\boldsymbol{\theta}^{1},\boldsymbol{\theta}^{2},\ldots,\boldsymbol{\theta}^{k}$, are sent to a centralized aggregation server, and an agglomerated model $\overline{\mathcal{M}}$ is made, and the model parameters $\boldsymbol{\Theta}^{}$ are exposed. The process is repeated until the model parameters are optimal.
• Figure 3: Architecture of the QNN used in this work. The process begins with data encoding through angle embedding, applying a series of $R_X(\theta_n)$ gates. These states are then processed by the QNN, which consists of alternating layers of $H$, $CNOT$, and parameterized rotation gates $R_Y(\theta_n)$. The unitary operations $U(\theta_i)$ further manipulate the quantum states. The output of the QNN is measured, and the result is used to calculate the MSE loss for model optimization. Parameter updates are performed iteratively using the Adam optimizer, adjusting the rotation angles to minimize the loss function, thereby refining the model with each iteration.
• Figure 4: The FedQNN framework where each Client/Device retains its Local Data, ensuring data privacy as no raw data is shared with the Central Server. Clients independently train a QNN model with their data, and only quantum model updates--comprising parameters or operations--are communicated to the central server, which functions as an Aggregation Point, synthesizing a global model from these updates. Secure communication protocols are employed for exchanging updates and preventing the exposure of individual data.
• Figure 5: Experimental setup and tool flow for conducting the experiments: The process begins with datasets fed into the QNN within the FedQNN framework. The QNN, implemented using the PennyLane framework, undergoes optimization through iterative training, with the number of iterations indicated for the Pennylane simulator and IBM QPUs. Post-training, the model's performance is assessed using standard binary classification metrics, including precision, recall, accuracy, and the F1 score.