QFed: Parameter-Compact Quantum-Classical Federated Learning

TL;DR

The paper tackles privacy-preserving distributed learning under data heterogeneity by introducing QFed, which embeds Quantum-Train to compress local classical models in a federated setting. QT uses a small quantum neural component to generate parameters for a larger classical model, enabling polylogarithmic reductions in trainable parameters while keeping inference classical. On FashionMNIST, QFed achieves a 77.6% reduction in parameter count (from $M=6690$ to $M=1497$) with accuracy close to non-quantum baselines, demonstrating practical gains in communication and computation in edge Federated Learning. The work highlights a viable hybrid quantum-classical path for scalable, privacy-preserving FL, leveraging cloud-based quantum resources without on-device quantum hardware, and discusses current limitations and future hardware-driven improvements.

Abstract

Organizations and enterprises across domains such as healthcare, finance, and scientific research are increasingly required to extract collective intelligence from distributed, siloed datasets while adhering to strict privacy, regulatory, and sovereignty requirements. Federated Learning (FL) enables collaborative model building without sharing sensitive raw data, but faces growing challenges posed by statistical heterogeneity, system diversity, and the computational burden from complex models. This study examines the potential of quantum-assisted federated learning, which could cut the number of parameters in classical models by polylogarithmic factors and thus lessen training overhead. Accordingly, we introduce QFed, a quantum-enabled federated learning framework aimed at boosting computational efficiency across edge device networks. We evaluate the proposed framework using the widely adopted FashionMNIST dataset. Experimental results show that QFed achieves a 77.6% reduction in the parameter count of a VGG-like model while maintaining an accuracy comparable to classical approaches in a scalable environment. These results point to the potential of leveraging quantum computing within a federated learning context to strengthen FL capabilities of edge devices.

Figures (5)

• Figure 1: The figure illustrates three key aspects: (1) the overall FL workflow, showing the interaction between edge devices and the central server; (2) a detailed view of one edge device with its core components: classical ML model, QNN, and mapping model; and (3) the proposed federated architecture that integrates QT, with edge devices simulated using Docker containers.
• Figure 2: Accuracy and loss curves for centralized classical training without Quantum-Train.
• Figure 3: The confusion matrix of the centralized model trained using the Quantum-Train approach.
• Figure 4: Federated global model performance using QT, accuracy and loss across different clients while communication rounds=70 and local epochs=10, number of clients/edges are tested with 5, 10, 20, 30, 40, 50, and 60.
• Figure 5: The global and local models' accuracies with different training settings, showing only the first 4 clients and the server.