Source-linked AI summary
Federated Quantum Machine Learning
Samuel Yen-Chi Chen, Shinjae Yoo
TL;DR
The paper addresses how to distribute QML computation across limited quantum machines while reducing privacy exposure, against a background in which federated QML had not yet been explored. It proposes federated training for hybrid quantum-classical classifiers, and reports that testing accuracy is not sacrificed while training converges quickly compared with non-federated training.
Problem
The paper targets the challenges of distributing computation across NISQ computers with limited capabilities and addressing privacy concerns in large-scale machine learning.
Method
The paper trains hybrid quantum-classical classifiers in a federated manner, using a central node to aggregate client parameters and distribute the global model.
Results
Federated training does not sacrifice testing accuracy and the model converges quickly compared with non-federated training.
Takeaways & Limitations
The framework supports privacy-preserving training and distributes computational loads across an array of NISQ computers.
Abstract
from arXiv · showhide
Distributed training across several quantum computers could significantly improve the training time and if we could share the learned model, not the data, it could potentially improve the data privacy as the training would happen where the data is located. However, to the best of our knowledge, no work has been done in quantum machine learning (QML) in federation setting yet. In this work, we present the federated training on hybrid quantum-classical machine learning models although our framework could be generalized to pure quantum machine learning model. Specifically, we consider the quantum neural network (QNN) coupled with classical pre-trained convolutional model. Our distributed federated learning scheme demonstrated almost the same level of trained model accuracies and yet significantly faster distributed training. It demonstrates a promising future research direction for scaling and privacy aspects.
I. INTRODUCTION
Quantum machine learning combines emerging quantum hardware with data-driven learning, but NISQ limitations and privacy concerns constrain practical deployment. The paper addresses these challenges through federated training of hybrid quantum-classical classifiers, reporting preserved testing accuracy and faster convergence than non-federated training.
- Motivation: NISQ devices cannot faithfully implement quantum circuits with many qubits or long circuit depth because they lack quantum error correction and suffer device noise.These constraints make applications requiring moderate quantum resources highly challenging.
- Motivation: Machine learning models require large datasets, including sensitive personal information such as medical records, browsing habits, and mobile-device data.Unauthorized access, compromised communication channels, and malicious cloud infrastructure create privacy concerns.
- Federated learning: Federated learning keeps training on users’ devices and exchanges models with the cloud rather than directly uploading recordings or other local data.This decentralized architecture is presented as a solution to privacy risks in large-scale machine learning.
- Contribution: The paper proposes federated training for hybrid quantum-classical classifiers to distribute computational tasks across quantum machines while addressing privacy concerns.The framework combines quantum machine learning with federated learning in the NISQ era.
- Results and implications: Federated training does not decrease testing accuracy and converges quickly compared with non-federated training.The authors position the approach as supporting secure QML infrastructure and better utilization of available NISQ devices.
II. FEDERATED MACHINE LEARNING
Federated quantum machine learning uses a central node to coordinate local training on quantum computers or simulators while keeping data distributed. Client circuit parameters are aggregated into a global model and shared back to clients.
- Architecture: Federated learning consists of a central node and several client nodes, with the central node holding and distributing the global model.Clients train locally using their own portions of the data, which are generally small.
- Quantum federated training: In the proposed framework, client nodes are quantum computers or quantum simulators whose circuit parameters are trained using a hybrid quantum-classical method.The central node receives trained parameters, aggregates them, and generates a new global model.
- Aggregation: The work aggregates client models by taking their mean after local training.The paper notes that multiple aggregation methods exist but selects the mean for this framework.
III. VARIATIONAL QUANTUM CIRCUITS
Variational quantum circuits encode classical data, apply learnable quantum operations, and measure outputs for further processing. Their architecture combines encoding, variational, and measurement components, with capacity constrained by available hardware.
- E(x) encodes classical data into a quantum state, while W(φ) is the learnable circuit block analogous to neural-network weights.Measured classical outputs can be processed by additional classical or quantum components.
- VQCs or QNNs combine an encoder, a variational layer, and quantum measurement.The encoder prepares quantum states, the variational layer applies trainable operations, and measurement returns classical values.
- The variational layer entangles neighboring qubits with CNOT gates and uses single-qubit unitary gates containing trainable parameters α, β, and γ.The parameters labeled α_i, β_i, and γ_i are optimized iteratively.
- Quantum measurements output Pauli-Z expectation values from designated qubits, with adjustable qubit and measurement counts.The reported classifier uses four qubits and measures the first two for binary classification.
- Circuit capacity depends on available quantum hardware or simulation software, and the grouped circuit block is repeated twice in this work.The architecture can repeat the grouped block to increase the number of parameters.
A. Quantum Encoder
The quantum encoder transforms classical values into quantum states using rotation angles derived from the inputs. The encoding must remain compact because NISQ hardware limits qubit counts and circuit depth.
- The encoding scheme determines how a classical vector is transformed into a quantum state and affects hardware implementation efficiency.It is also discussed in relation to potential quantum advantages.
- NISQ limitations require encoding classical values with few qubits and without too many quantum operations.The number of qubits and circuit depth are limited on near-term devices.
- An N-qubit state is represented by amplitudes c_{q1,...,qN} over basis states indexed by q_i ∈ {0, 1}.The amplitudes are complex, and their squared magnitudes determine measurement probabilities whose total is one.
- Variational encoding uses input values or their transformations as rotation angles for quantum rotation gates.The reported encoding uses Ry and Rz gates with arctan-based transformations.
B. Quantum Gradients
The hybrid quantum-classical model is trained end to end with backpropagation, using the parameter-shift method to calculate analytical gradients for quantum circuits.
- Hybrid quantum-classical models can be trained end to end using the backpropagation method used for deep neural networks.
- The parameter-shift method calculates analytical gradients for quantum-circuit functions without finite differences.It uses expectation values of quantum observables to obtain quantum gradients.
IV. HYBRID QUANTUM-CLASSICAL TRANSFER LEARNING
The paper combines a pre-trained classical VGG16 model with a four-qubit variational circuit in a federated training setting. On Cats vs Dogs, federated and non-federated models reached comparable testing performance, while training loss fluctuated across clients but converged.
- IV. HYBRID QUANTUM-CLASSICAL TRANSFER LEARNING: NISQ devices limit qubit counts and circuit depth, making suitable near-term quantum model architectures non-trivial to design.The paper addresses this setting with hybrid quantum-classical transfer learning.
- IV. HYBRID QUANTUM-CLASSICAL TRANSFER LEARNING: A central node aggregates circuit parameters from local machines into a global parameter and distributes it back to them.Local machines train on distinct data portions for E epochs with batch size B.
- A. Cats vs Dogs: The experiment uses 100 local machines, 23,000 training points, 2,000 testing points, and five randomly selected machines per round.Each local machine has 230 training points, and selected machines train for 1, 2, or 4 epochs.
- A. Cats vs Dogs: Collected model averaging aggregates local models, and the global model is shared with local machines after each round.The study performs 100 training rounds and compares federated with non-federated training using the same architecture and dataset.
- A. Cats vs Dogs: The hybrid model pairs a pre-trained VGG16 model with a 4-qubit VQC receiving 4-dimensional compressed vectors.The original VGG16 classifier is replaced to fit the VQC input dimension; the repeated circuit contains 24 parameters.
- A. Cats vs Dogs: Testing accuracies and testing loss reach levels comparable to non-federated training, regardless of the number of local training epochs.Training loss fluctuates because different clients are selected in different rounds, but it converges after 100 rounds.
- A. Cats vs Dogs: A single local training epoch is sufficient to train a well-performing model in the reported experiment.
B. CIFAR (Planes vs Cars)
On CIFAR-10 Planes vs Cars, federated hybrid quantum-classical training reaches testing performance comparable to non-federated training across local-epoch settings. One local epoch is reported as sufficient for a well-performing model, with updates formed from five clients per round.
- The experiment uses CIFAR-10 images with dimensions of 32 × 32 and presents federated results in Figure 8.
- The global model is evaluated on 2,000 centralized testing points, while 10,000 training points are distributed across local machines.
- Each federated round randomly selects five of 100 local machines, each training for 1, 2, or 4 epochs before central aggregation.
- Testing accuracies and testing loss reach comparable levels to non-federated training across local training-epoch settings.
- A single local training epoch is reported as sufficient to train a well-performing model.
A. Integration with Other Private-Preserving Protocols
The paper discusses privacy-preserving protocols that could complement federated quantum learning, while identifying model exchange and aggregation as security vulnerabilities. Differential privacy and secure multi-party computation are presented as possible protections, and advanced aggregation remains future work.
- Exchanged model parameters may be attacked, and leaked parameters can be used to deduce training data or training membership.
- These inference risks raise concerns when QML models process private and sensitive data.
- Differential privacy could allow trained models to be shared while retaining private information in the training data.
- Secure multi-party computation is proposed as another direction for increasing security in decentralized computing.
- Implementing advanced aggregation protocols for quantum machine learning is identified as an interesting direction for future work.
C. Decentralization
The paper positions decentralized federated quantum learning as a future extension beyond its central-node framework. Blockchain, peer-to-peer, and gossip approaches could remove central coordination, but their quantum-regime efficiency and capabilities remain unstudied here.
- The presented research is a proof-of-concept for federated training on quantum machine learning.
- The current federated scheme uses a central node to receive local models, aggregate them, and distribute the aggregated model.
- A central node can be vulnerable to malicious attacks, while variable client bandwidth can disrupt synchronization.
- Blockchain-enabled and peer-to-peer federated schemes are discussed as approaches that can remove the need for a central node.
- Under gossip learning, network nodes exchange and aggregate models directly without a central node.
- The efficiency and capabilities of blockchain-based and gossip learning schemes in the quantum regime are left for future work.
D. Other Quantum Machine Learning Models
The proposed federated framework uses hybrid quantum-classical transfer learning with a pre-trained classical feature extractor, but is described as extensible to broader quantum models and applications. Possible directions include QCNNs, speech recognition, healthcare, and finance.
- The considered architecture combines a pre-trained classical feature extractor with a quantum component in hybrid quantum-classical transfer learning.
- The framework can be extended beyond transfer learning and could be investigated with more generic tensor-network and quantum-circuit architectures.
- Federated learning for quantum convolutional neural networks is suggested when larger-scale quantum simulators or real quantum computers become available.
- Limited qubit availability in current quantum computers and simulators motivates the use of the transfer-learning structure.
- Potential applications include decentralizing quantum-enhanced speech recognition and privacy-preserving healthcare and finance models.
- The work provides a federated framework for hybrid quantum-classical classifiers across NISQ computers without sacrificing testing accuracy.