[Paper] QFed: Parameter-Compact Quantum-Classical Federated Learning

Source: arXiv - 2601.09809v1

Overview

The paper “QFed: Parameter‑Compact Quantum‑Classical Federated Learning” explores how quantum computing can be woven into federated learning (FL) pipelines to dramatically shrink model size without sacrificing accuracy. By leveraging quantum‑enhanced layers, the authors demonstrate a 77 % reduction in parameters for a VGG‑style network on the FashionMNIST benchmark, opening the door for more lightweight FL on edge devices that are constrained by memory, compute, and power.

Key Contributions

• Quantum‑augmented FL framework (QFed): Introduces a hybrid architecture where selected layers are implemented as quantum circuits, reducing the classical parameter count by polylogarithmic factors.
• Parameter‑compact design: Shows that a VGG‑like model can be compressed from ~13 M to ~3 M trainable parameters while preserving classification performance.
• Scalable experimental validation: Implements QFed on a realistic federated simulation (multiple clients, non‑IID data splits) using the FashionMNIST dataset and reports comparable test accuracy to the fully classical baseline.
• Practical performance analysis: Quantifies the trade‑off between quantum circuit depth, communication overhead, and overall training time, highlighting scenarios where QFed yields net efficiency gains.

Methodology

1. Hybrid Model Construction – The authors replace a subset of convolutional/fully‑connected layers with variational quantum circuits (VQCs). Each VQC acts as a parameterized feature extractor; its trainable angles serve as the “quantum parameters.”
2. Federated Training Loop – Standard FL steps (client‑side local training, server‑side aggregation) are retained. Clients download the current hybrid model, perform a few local epochs on their private data, and send back updated quantum and classical weights.
3. Parameter Count Analysis – By mapping a classical weight matrix of size (N \times N) to a quantum circuit with (\log N) qubits, the number of trainable parameters scales as (\mathcal{O}(\log^2 N)) instead of (\mathcal{O}(N^2)).
4. Simulation Setup
   • Dataset: FashionMNIST (10 classes, 60 k training images).
   • Clients: 10 simulated edge devices, each receiving a non‑IID slice of the data.
   • Baseline: Classical VGG‑like network with ~13 M parameters.
   • Metrics: Parameter count, test accuracy, communication volume, and wall‑clock training time.

Results & Findings

*Measured on a simulated edge device with a CPU‑only backend; quantum circuit execution was emulated using a statevector simulator.

Interpretation: The hybrid model retains virtually the same predictive power while slashing the number of parameters and slightly lowering communication payloads. The modest speed‑up in local compute time suggests that, when real quantum hardware with low latency becomes available, the gains could be even larger.

Practical Implications

• Edge‑friendly FL: Devices such as smartphones, IoT sensors, or wearables often lack the memory to host deep CNNs. QFed’s compact representation makes it feasible to run richer models locally, reducing reliance on heavyweight server inference.
• Bandwidth‑constrained deployments: Fewer parameters mean smaller model updates, which translates to lower network usage—critical for remote or mobile scenarios with limited connectivity.
• Energy efficiency: Smaller models consume less power during training and inference, extending battery life for edge hardware.
• Roadmap for quantum‑enhanced AI services: Cloud providers could expose quantum‑accelerated model components as a service, letting FL orchestrators swap in quantum layers where they deliver the biggest parameter savings.

Limitations & Future Work

• Quantum simulation overhead: The experiments used a classical simulator; real quantum hardware introduces latency, noise, and qubit‑count limits that could offset the theoretical parameter savings.
• Scope of model compression: Only a portion of the network was quantum‑enabled; extending the approach to deeper or more complex architectures may encounter scalability bottlenecks.
• Security & privacy analysis: While FL already protects raw data, the impact of quantum‑based weight updates on differential privacy guarantees remains unexplored.
• Future directions: The authors suggest (1) testing on actual NISQ devices, (2) exploring adaptive layer selection (which layers to quantum‑ify per client), and (3) integrating quantum‑aware compression techniques such as quantum‑aware pruning or quantization.

Authors

• Samar Abdelghani
• Soumaya Cherkaoui

Paper Information

• arXiv ID: 2601.09809v1
• Categories: cs.LG, cs.AI, cs.DC
• Published: January 14, 2026
• PDF: Download PDF