Quantum Privacy Aggregation of Teacher Ensembles (QPATE) for Privacy-preserving Quantum Machine Learning

TL;DR

This work addresses privacy in quantum machine learning by adapting Privacy-Aggregation of Teacher Ensembles (PATE) to ensembles of quantum neural networks, yielding Quantum PATE (qPATE). A hybrid quantum-classical classifier is trained where noisy, privacy-preserving teacher labels guide a quantum student, with privacy loss bounded via a momentum-accountant and DP mechanisms. Empirical results on MNIST binary classification show that quantum PATE can outperform classical PATE at low privacy budgets (small $\epsilon$), while maintaining competitive accuracy as $\epsilon$ increases. The study demonstrates a practical, privacy-preserving pathway for quantum classifiers and points to potential quantum advantages in privacy-constrained QML, with future work extending to larger-scale vision tasks.

Abstract

The utility of machine learning has rapidly expanded in the last two decades and presents an ethical challenge. Papernot et. al. developed a technique, known as Private Aggregation of Teacher Ensembles (PATE) to enable federated learning in which multiple teacher models are trained on disjoint datasets. This study is the first to apply PATE to an ensemble of quantum neural networks (QNN) to pave a new way of ensuring privacy in quantum machine learning (QML) models.

Figures (4)

• Figure 1: Classical PATE network architecture. Classical PATE uses four convolution blocks.
• Figure 2: Quantum PATE network architecture. Quantum PATE uses two convolution blocks with the additional VQC block.
• Figure 3: VQC block for MNIST classification. The VQC block encodes latent embeddings from convolution blocks within quantum PATE into quantum states represented by $10$-qubits. $U(\mathbf{x})$ denotes the quantum algorithm for angle encoding. $\phi_{i}$, $\theta_{i}$, and $\omega_{i}$ are the parameters to optimize. The dashed box denotes one subcircuit of the VQC block that is repeated two times. The dial to the far right represents that the circuit has two outputs. The expectation of $\sigma_z$ is measured on two qubits.
• Figure 4: Accuracy vs. Epsilon Plot for 4 Teachers in classical PATE and quantum PATE. We averaged the results of 10 experiments, and the error bar denotes the standard deviation. (A) is the result of 1 epoch training, (B) is the result of 10 epoch training, (C) is the result of 20 epoch training.