Practical quantum federated learning and its experimental demonstration

TL;DR

This work proposes a practical quantum federated learning framework on quantum networks, utilizing distributed quantum secret keys to protect local model updates and enable secure aggregation with information-theoretic security and provides critical insights for building scalable, efficient, and quantum-secure machine learning systems for the coming quantum internet era.

Abstract

Federated learning is essential for decentralized, privacy-preserving model training in the data-driven era. Quantum-enhanced federated learning leverages quantum resources to address privacy and scalability challenges, offering security and efficiency advantages beyond classical methods. However, practical and scalable frameworks addressing privacy concerns in the quantum computing era remain undeveloped. Here, we propose a practical quantum federated learning framework on quantum networks, utilizing distributed quantum secret keys to protect local model updates and enable secure aggregation with information-theoretic security. We experimentally validate our framework on a 4-client quantum network with a scalable structure. Extensive numerical experiments on both quantum and classical datasets show that adding a quantum client significantly enhances the trained global model's ability to classify multipartite entangled and non-stabilizer quantum datasets. Simulations further demonstrate scalability to 200 clients with classical models trained on the MNIST dataset, reducing communication costs by $75\%$ through advanced model compression techniques and achieving rapid training convergence. Our work provides critical insights for building scalable, efficient, and quantum-secure machine learning systems for the coming quantum internet era.

Figures (5)

• Figure 1: Schematic of the QuNetQFL Framework. (a) The QuNetQFL framework employs a fully connected quantum network among $K$ clients, where each client uses QKD to securely exchange quantum secure keys with others, enabling pairwise masking of local model updates. Only these masked updates are sent to the server through classical channels, preserving client privacy. The server aggregates the updates to form an unmasked global model parameter, which is then redistributed to all clients. (b) Masked secure aggregation in a four-client setup. Here, $\mathbf{x}_i$ represents the raw local updates and $\mathbf{y}_i$ the masked updates. The server simply sums the masked updates to obtain the decrypted global model update. (c) and (d) Two types of QNN representations with hardware-efficient ansatzes used in the study.
• Figure 2: Experimental setup of the quantum network. At Eve’s site, a continuous-wave laser source is employed with two intensity modulators (IM) and a variable optical attenuator (VOA) to produce weak coherent pulses. Four clients (Alice, Bob, Charlie, and David) are interconnected in the Sagnac loop. When any pair of clients need to establish secret keys, Eve injects the pulses into the loop via a circulator (Cir) and a $50:50$ beam splitter (BS). Upon reaching the designated clients, phase modulators (PMs) are employed to add phase to the pulses. After modulation, the two pulse trains interfere at Eve’s BS, where interference results are detected by two superconducting nanowire single-photon detectors ($D_1$ and $D_2$). This structure includes 1-km single-mode fibers (SMFs) between Alice and Eve, Bob and Eve, Alice and Charlie, and Bob and David, while a 2-km SMF connects Charlie and David. Attenuators (Att) are placed between Alice and Eve, and between Bob and Eve, to adjust channel losses. Polarization controllers (PCs) are used to align polarization in the loop.
• Figure 3: Performance of QuNetQFL on quantum state classification using QNN. (a) Entanglement classification: final test accuracies are $88.5\%$ with three clients, $91.5\%$ with four clients, and $93\%$ for the benchmark. (b) Nonstabilizerness classification: final test accuracies are $95.8\%$ with three clients, $98.3\%$ with four clients, and $100\%$ for the benchmark. These two tasks used $16$-bit quantization ($q = 16$) over $200$ and $160$ communication rounds, respectively. Adding a single client notably improves global model accuracy, bringing it closer to the benchmark.
• Figure 4: Evaluation of QuNetQFL on the MNIST classification task using QNN. (a) IID setting showing test accuracy and loss across 200 communication rounds for client combinations $\{3,6\}$, $\{0,1\}$, $\{3,5\}$, and $\{3,9\}$ with $16$-bit quantization. (b) Non-IID setting with identical client configurations as (a) for comparative analysis. (c) Final test accuracy comparison between IID and non-IID settings, with annotated accuracy differences illustrating QuNetQFL's robustness across data distributions. (d) Non-IID client data distribution proportions for classes $\{3,6\}$ (same for other cases), highlighting dataset heterogeneity.
• Figure : Protocol of QuNetQFL