Quantum Federated Learning Experiments in the Cloud with Data Encoding

TL;DR

The work tackles deploying Quantum Federated Learning (QFL) on cloud infrastructure using Qiskit, focusing on encoding classical data into quantum states and federated aggregation to train quantum neural networks without sharing raw data. It proposes an encoding-driven QFL architecture, leveraging a feature map, a parameterized quantum circuit, and central aggregation, and evaluates it on a genomic dataset using IBM Cloud simulators. The study compares three aggregation schemes (Simple Averaging, Weighted Averaging, Best Pick) and demonstrates that amplitude encoding enables efficient handling of high-dimensional genomic data under a qubit budget. The results suggest encoding choices and aggregation strategies can improve global model performance while preserving privacy, and they outline future work toward real hardware deployment and standardization of QFL in cloud ecosystems.

Abstract

Quantum Federated Learning (QFL) is an emerging concept that aims to unfold federated learning (FL) over quantum networks, enabling collaborative quantum model training along with local data privacy. We explore the challenges of deploying QFL on cloud platforms, emphasizing quantum intricacies and platform limitations. The proposed data-encoding-driven QFL, with a proof of concept (GitHub Open Source) using genomic data sets on quantum simulators, shows promising results.

Figures (9)

• Figure 1: A high level view of local learning in the proposed QFL Process consisting of several key components. The Feature Map ingests input data and encodes it into a quantum state. Following this, the Ansatz comes into play as a parameterized quantum circuit, with its parameters being iteratively fed by the Optimizer--optimization objective function is driven by the outcomes from the Sampler.
• Figure 2: QFL Process. Each client in the setup trains models locally and only shares its parameters with a central server (step 1). The server aggregates these to enhance the global model(step 2), then circulates the updated parameters back to clients(step 3), preserving data privacy as client data stays local.
• Figure 3: Details of the QFL framework. Each client possesses unique data (V1, V2, V3), which undergoes transformation into quantum states using a Feature Map of size 255. These quantum states are then processed by a parameterized quantum circuit (Ansatz) with a depth of 3 and 32 weights. Local training using these quantum components is performed using the Qiskit Simulator. After training, the updated weights from the Ansatz are sent to the server for aggregation, where parameters from all clients are combined, and finally, the aggregated global weights are returned to the clients for local model updates.
• Figure 4: Comparision of the evolution of Top-1 Accuracy over epochs for the global model and clients models using the averaging technique.
• Figure 5: Evolution of Top-1 Accuracy and Loss over epochs for clients using the averaging technique.