Abstract
While the majority of focus in quantum computing has so far been on monolithic quantum systems, quantum communication networks and the quantum internet in particular are increasingly receiving attention from researchers and industry alike. The quantum internet may allow a plethora of applications such as distributed or blind quantum computing, though research still is at an early stage, both for its physical implementation as well as algorithms; thus suitable applications are an open research question. We evaluate a potential application for the quantum internet, namely quantum federated learning. We run experiments under different settings in various scenarios (e.g. network constraints) using several datasets from different domains and show that (1) quantum federated learning is a valid alternative for regular training and (2) network topology and nature of training are crucial considerations as they may drastically influence the models performance. The results indicate that more comprehensive research is required to optimally deploy quantum federated learning on a potential quantum internet.
Introduction
As quantum computing transitions from isolated, monolithic systems toward a decentralized "Quantum Internet," a fundamental question emerges: what are the practical applications of such a network? While distributed quantum computing and quantum key distribution are frequently discussed, Quantum Machine Learning (QML) presents a compelling alternative.
A significant challenge in realizing a global quantum network is hardware heterogeneity. Future networks will likely comprise diverse nodes with varying architectures and qubit capacities. This research explores Quantum Federated Learning (QFL) as a solution, enabling decentralized training across heterogeneous devices while preserving data privacy. We specifically investigate how different network topologies and communication constraints influence the learning process.
Methodology
The study utilizes Variational Quantum Circuits (VQCs)—hybrid algorithms where quantum parameters are optimized via classical routines—as the core machine learning models. To address the challenge of training across a decentralized network, two primary communication architectures were evaluated:
- Star Topology: A centralized "Global Model" server aggregates weights from multiple clients, performs updates, and redistributes the refined parameters.
- Ring Topology: Clients are arranged in a sequential loop, where each node trains on local data before passing its weights directly to the next neighbor.
A critical technical hurdle in these experiments was node heterogeneity. For instance, if a client with 4 qubits communicates with one having 6, their weight vectors differ in size. The authors implemented an ad-hoc mapping strategy: superfluous weights are discarded when moving to smaller devices, while missing weights are filled using the target device's existing parameters.
The models were tested on three binary classification datasets: the synthetic Moons dataset for non-linear classification, FashionMNIST (clothing images), and PneumoniaMNIST (medical X-rays). Images were preprocessed to accommodate varying qubit counts (4 to 10 qubits).
Results and Discussion
The experiments demonstrate that QFL is a viable alternative to centralized training, with performance often matching the "Quantum Baseline." However, the choice of network topology introduces significant trade-offs:
- Performance vs. Stability: The Star topology generally achieved higher training accuracy but exhibited significant volatility during the optimization process. In contrast, the Ring topology provided a smoother convergence curve and demonstrated superior performance on unseen test data.
- Resilience to Heterogeneity: Despite the simplicity of the weight-mapping strategy (truncation and padding), the models learned effectively across diverse hardware. This suggests that a functional Quantum Internet does not strictly require uniform hardware to support robust machine learning applications.
- Comparison with Classical Baselines: While classical models currently maintain an edge due to their massive parameter counts, QFL models performed comparably to their centralized quantum counterparts, indicating that the federated approach does not inherently degrade quantum performance.
Conclusion
This research confirms that Quantum Federated Learning is a promising application for future quantum networks. While decentralized training is effective even with heterogeneous hardware, the specific network topology plays a crucial role in model stability and generalization.
Future work must address the impact of real-world quantum noise and decoherence, which were not fully modeled in these simulations. Additionally, exploring quantum-native communication methods—such as exchanging weights via quantum teleportation—and developing more sophisticated weight aggregation algorithms will be essential for scaling QFL to larger, more complex quantum internets.
Summary updated