Source-linked AI summary
Quantum Federated Learning with Quantum Data
Mahdi Chehimi, Walid Saad
TL;DR
Centralized QML does not scale well to distributed quantum networks, and transferring fragile quantum data is difficult. This paper proposes quantum federated learning that shares circuit parameters rather than qubits, and reports performance comparable to or sometimes better than centralized QML while handling IID and non-IID data.
Problem
Existing QML models are centralized, while fragile qubits are difficult to transfer, leaving no comprehensive framework for distributed learning by quantum clients on quantum data.
Source: §1.2 · p. 2, §1.3 · p. 3Method
The framework uses QCNNs on quantum data, generates a quantum federated dataset, and aggregates clients’ learnable circuit parameters with federated averaging.
Source: §4.1 · p. 7, §6 · p. 11Results
The framework achieves comparable and sometimes superior performance to centralized purely quantum ML and handles IID and non-IID quantum data.
Source: §1.4 · p. 4Takeaways & Limitations
The results support applying classical federated-learning algorithms to decentralize learning in purely quantum QML applications.
Source: §1.4 · p. 4NISQ hardware’s small qubit capacity and heterogeneous device capabilities make practical deployment of large-scale QFL challenging.
Source: §5 · p. 10
Abstract
from arXiv · showhide
Quantum machine learning (QML) has emerged as a promising field that leans on the developments in quantum computing to explore large complex machine learning problems. Recently, some purely quantum machine learning models were proposed such as the quantum convolutional neural networks (QCNN) to perform classification on quantum data. However, all of the existing QML models rely on centralized solutions that cannot scale well for large-scale and distributed quantum networks. Hence, it is apropos to consider more practical quantum federated learning (QFL) solutions tailored towards emerging quantum network architectures. Indeed, developing QFL frameworks for quantum networks is critical given the fragile nature of computing qubits and the difficulty of transferring them. On top of its practical momentousness, QFL allows for distributed quantum learning by leveraging existing wireless communication infrastructure. This paper proposes the first fully quantum federated learning framework that can operate over quantum data and, thus, share the learning of quantum circuit parameters in a decentralized manner. First, given the lack of existing quantum federated datasets in the literature, the proposed framework begins by generating the first quantum federated dataset, with a hierarchical data format, for distributed quantum networks. Then, clients sharing QCNN models are fed with the quantum data to perform a classification task. Subsequently, the server aggregates the learnable quantum circuit parameters from clients and performs federated averaging. Extensive experiments are conducted to evaluate and validate the effectiveness of the proposed QFL solution. This work is the first to combine Google's TensorFlow Federated and TensorFlow Quantum in a practical implementation.
1 Introduction
The paper proposes quantum federated learning for collaborative learning on quantum data without transferring qubits. It motivates the framework by the limits of centralized QML and the difficulty of transmitting fragile quantum data, and situates it against prior decentralized approaches that do not address this setting.
1 Introduction
- The introduction situates QML within advances in quantum computing and describes its expansion into applications including communication networks, AI, and machine learning.It notes prior claims about quantum-computing speed and emerging applications.Source: §1 · pp. 1–2
1.1 Quantum Machine Learning
- Quantum many-body problems challenge classical and hybrid ML because of complex states, exponentially large Hilbert spaces, and difficult quantum-state tomography, motivating purely quantum models such as QCNNs.The introduction describes parametrized quantum circuits as tunable quantum neural networks and notes prior QCNN classification work.Source: §1.1 · p. 2
1.2 Quantum Communication Networks
- Fragile qubits make quantum-data transmission difficult, while existing QML models are centralized; the paper therefore targets collaborative quantum learning over classical wireless infrastructure without transferring quantum data.The introduction frames this as a response to the practical difficulty of transmitting quantum data and the lack of a comprehensive framework for purely quantum clients.Source: §1.2 · pp. 2–3, §1.3 · p. 3
1.3 Problem Statement
- The paper poses practical questions about generating quantum federated datasets, learning circuit parameters with classical FL algorithms, hardware constraints, and clients’ differing data distributions.These questions define the scope of the problem statement, including IID and non-IID client data.Source: §1.3 · p. 3
1.4 Contributions
- The proposed framework decentralizes QCNN circuit-parameter learning across quantum clients and aggregates parameters at a server, with experiments showing comparable or sometimes superior performance to centralized QML and support for IID and non-IID data.The authors also identify applying classical federated-learning algorithms to purely quantum QML applications as a key finding.Source: §1.4 · pp. 3–4
1.5 Related Works
- Earlier decentralized QML work did not provide a practical framework for large-scale networks of quantum clients learning from purely quantum data.The introduction distinguishes ’s single-client classical task, ’s conceptual classical-user architecture, ’s quantum-server feature extraction from classical data, and ’s federated transfer-learning approach using classical data; it identifies the purely quantum-data setting as unaddressed.Source: §1.5 · p. 4
2 Problem Setup
The setup distributes quantum-data classification across clients that train a shared QCNN, while a server coordinates model parameters over classical wireless infrastructure. Clients optimize locally, and the server aggregates their parameters for the next training round.
2 Problem Setup
- Clients train a shared QCNN on locally generated quantum data for binary classification, while the server coordinates the distributed learning process.The framework uses classical wireless communication infrastructure for decentralized quantum learning.Source: §2 · pp. 4, 6
- QCNN predictions use convolution and pooling layers followed by a fully connected layer and measurements, with pooling reducing system size and introducing nonlinearities.The learnable client parameters are entries of the circuit’s unitaries, and training minimizes mean squared error.Source: §2 · p. 5
- Each client uses local quantum data to update its model parameters, then sends the updated parameters to the server for aggregation and redistribution.The server applies Federated Averaging, weights client contributions, and repeats the process until convergence.Source: §2 · p. 6
3 Quantum Federated Dataset
The section introduces a hierarchical-format federated dataset of purely quantum data, built from quantum cluster states and organized into client datasets for distributed learning.
3 Quantum Federated Dataset
- The proposed framework generates a first quantum federated dataset for distributed quantum-network learning, using purely quantum data in a hierarchical data format.The dataset addresses the stated lack of quantum federated datasets in the literature.Source: §3 · p. 6
3.1 Quantum Data
- The dataset uses quantum cluster states with excitations labeled according to qubit rotations, a data type associated with quantum sensing and distributed quantum computing networks.The authors also note the states’ ability to teleport quantum states between clients communicating through a quantum channel.Source: §3.1 · p. 6
3.2.1 Generating Single Client Data
- Single-client data are generated on a 1 × 8 qubit grid as cluster-state circuits, with Rx rotations used to represent excitations and a sufficiently large rotation determining the label.The circuits use Hadamard and Controlled-Z gates; the text describes excited states as labeled 1 and unexcited states as labeled 0.Source: §3.2.1 · pp. 6–7
3.2.2 Generating Federated Data
- The HDF5 federated dataset contains 30 clients, each with 160 serialized quantum-data inputs and integer labels; the authors also generate non-IID client datasets for comparison with the IID case.The number of clients can be varied.Source: §3.2.2 · p. 7
4 Experiments
Experiments test QFL across client counts, dataset sizes, optimizers, and data distributions. Testing accuracy generally improves with more clients, federation outperforms centralized learning in the tested dataset-size comparison, and performance is similar on IID and non-IID data.
4 Experiments
- The experiments evaluate the proposed QFL framework’s applicability and effectiveness.The section first details the experimental setup, then reports experiments.Source: §4 · p. 7
4.1 Experimental Setup
- The eight-qubit setup uses a QCNN with three convolution-pooling pairs and 64 learnable parameters; client optimizers are compared while server SGD is fixed for averaging.The setup uses TFQ and TFF in Google Colaboratory on CPU nodes; all available clients are assumed to train, with some clients’ data reserved for testing.Source: §4.1 · p. 7
4.2.1 Impact of Number of Clients
- Testing accuracy generally increases as participating clients grow, without training overfitting; the six-client case achieves lower accuracy.The comparison fixes each client’s dataset at 160 samples; the authors attribute the six-client result to too few clients for efficient learning.Source: §4.2.1 · p. 8
4.2.2 Impact of the Size of Clients’ Data Samples
- The 30-client federated framework outperforms the centralized case, while increasing client dataset size does not necessarily improve testing accuracy.With a shallow QCNN and 64 trainable parameters, more data may slightly increase or decrease accuracy once each client has enough data.Source: §4.2.2 · p. 8
4.2.3 Impact of Optimizers and Learning Rates
- Adam at a 0.02 learning rate converges to higher testing accuracy after a few epochs than SGD at 0.02, which reaches around 96.5%.RMSprop at 0.002 converges more slowly and to lower accuracy; Adam at 0.2 cannot learn, while Adam at 0.002 yields a smoother curve at the expense of lower accuracy.Source: §4.2.3 · pp. 8–9
4.2.4 Impact of Non-IID Data
- The QFL framework achieves similar performance on IID and non-IID quantum federated datasets.The comparison reports testing accuracy and MSE loss for 30 clients, with 25 training clients and 5 testing clients, each holding 160 samples.Source: §4.2.4 · p. 9, Table 1 · p. 10
4.2.5 Error Bars
- Random-seed errors across repeated runs are below 5% for training and testing performance metrics.The tested network has 30 clients with 160 samples each, and 5 clients’ data is used for testing.Source: §4.2.5 · p. 10, Figure 5 · p. 10
5 Challenges and Future Outlook
The framework faces practical deployment challenges from limited NISQ qubit capacity and heterogeneous quantum devices. The authors also outline opportunities in wireless integration and future work on optimization and security.
5 Challenges and Future Outlook
- NISQ hardware processes only a small number of qubits, complicating practical deployment of large-scale purely quantum QML models.The authors also note that differing device technologies and capabilities may make adoption of a unified QML model difficult.Source: §5 · p. 10
- The framework is proposed as a way to integrate quantum devices and quantum data with existing wireless networks.The authors anticipate applications and research problems in networking, quantum hardware, and wireless sensing, including optimization for networks with quantum and classical clients.Source: §5 · p. 10
- Future work could investigate encrypting classical learning parameters with quantum cryptographic schemes and integrating QFL with QKD.The authors frame QFL with QKD as an open research problem.Source: §5 · p. 10
6 Conclusion
The paper proposes a QFL framework for distributed learning over quantum data without sending qubits, using classical wireless networks. It introduces a quantum federated dataset and a TensorFlow Quantum–TensorFlow Federated implementation, with experiments validating federated averaging's effective behavior.
6 Conclusion
- The proposed framework supports scalable distributed quantum learning over quantum data without sending qubits, leveraging classical wireless networks.The authors describe the framework as a novel QFL approach.Source: §6 · p. 11
- The work generates the first quantum federated dataset in the literature and implements a combination of TensorFlow Quantum and TensorFlow Federated.The authors describe the implementation as unique.Source: §6 · p. 11
- Experiments validate the effective behavior of the proposed QFL framework using federated averaging.The authors report conducting extensive experiments to evaluate applicability and effectiveness.Source: §6 · p. 11