NAC-QFL: Noise Aware Clustered Quantum Federated Learning

TL;DR

NAC-QFL tackles the core bottlenecks of near-term quantum learning in distributed mobile environments by integrating noise-aware device selection with cluster-based quantum federated learning. It introduces a clustering workflow driven by entanglement-assisted channel capacity and a polynomial-time device selection routine that accounts for hardware noise via a calibrated N_eff model. The approach leverages circuit cutting to train smaller subcircuits on higher-fidelity devices and employs FedAvg-like aggregation weighted by cluster quantum volumes to improve convergence and reduce communication. Experimental results on MNIST/P-MNIST and noisy datasets demonstrate improved accuracy, faster convergence, and resilience to noisy channels, highlighting practical potential for noise-robust QFL on NISQ hardware.

Abstract

Recent advancements in quantum computing, alongside successful deployments of quantum communication, hold promises for revolutionizing mobile networks. While Quantum Machine Learning (QML) presents opportunities, it contends with challenges like noise in quantum devices and scalability. Furthermore, the high cost of quantum communication constrains the practical application of QML in real-world scenarios. This paper introduces a noise-aware clustered quantum federated learning system that addresses noise mitigation, limited quantum device capacity, and high quantum communication costs in distributed QML. It employs noise modelling and clustering to select devices with minimal noise and distribute QML tasks efficiently. Using circuit partitioning to deploy smaller models on low-noise devices and aggregating similar devices, the system enhances distributed QML performance and reduces communication costs. Leveraging circuit cutting, QML techniques are more effective for smaller circuit sizes and fidelity. We conduct experimental evaluations to assess the performance of the proposed system. Additionally, we introduce a noisy dataset for QML to demonstrate the impact of noise on proposed accuracy.

Figures (9)

• Figure 1: Effect of device topology and noise model on the circuit performance. Parts (a)-(b) show the effect of topology on three qubit GHZ circuit. Part (c) shows the effect of a noise model for the same topology seven-qubit IBM device. Part (d) shows the performance variation of five-qubit GHZ performance in different capacity IBM devices.
• Figure 2: Noisy Quantum Channel: Encoding quantum information and transmitting it over a quantum communication channel which adds noise and quantum information is recovered after decoding.
• Figure 3: Workflow for NAC-QFL system. Steps and are used for cluster creation and cluster head selection. Steps -are used for noise-aware device selection phase. Step initiates model distribution to the cluster heads. Steps -are for Distributed model training within a cluster Step is used for trained model upload to the central server and Step used for cluster wise aggregation.
• Figure 4: () 1: Clustering in NAC-QFL system
• Figure 5: () 2: Noise Aware Device Selection.

Theorems & Definitions (2)

• Theorem 1
• Corollary 1.1