Advances in Quantum Federated Learning

• Quantum Federated Learning is a decentralized framework that integrates quantum computing with federated model training to enhance privacy and efficiency.
• It leverages local parameterized quantum circuits and quantum natural gradient descent to optimize training while exchanging only classical model updates.
• The architecture incorporates advanced encryption, quantum-secure communication, and blockchain to safeguard data and improve robustness against noise and adversarial threats.

Quantum Federated Learning (QFL) is a framework that merges the decentralized, privacy-preserving paradigm of classical federated learning (FL) with the computational and representational strengths of quantum machine learning (QML). QFL enables distributed collaborative training of quantum models across multiple clients—each potentially endowed with quantum processors—without transferring local data, instead exchanging model updates derived from quantum circuits. This architecture is designed to overcome classical FL limitations in privacy, communication costs, and efficiency by leveraging quantum effects such as superposition, entanglement, and quantum-secured communication.

1. Fundamental Principles and Architectural Foundations

QFL combines the key features of FL—model parameter aggregation across clients, data privacy via local training, and robustness to non-IID data—with quantum-native computing methods. At the core of QFL, clients run parameterized quantum circuits (PQCs) or quantum neural networks (QNNs) on their local data, optimizing quantum gate parameters locally. Instead of exchanging quantum states—which are fragile and infeasible to transmit over classical wireless channels—only classical parameters describing the trained quantum circuits are shared (Chehimi et al., 2021). A server or decentralized aggregator then aggregates these classical updates, typically using weighted federated averaging, and distributes the updated model parameters back to clients for further training iterations.

Mathematically, the typical global update in QFL has the form:

$$\theta_{h+1}^{\mathrm{global}} = \sum_{k \in \mathcal{K}} w_k \cdot \theta_h^k$$

where $w_k$ are the aggregation weights (e.g., proportional to the clients’ data sizes), and $\theta_h^k$ are the PQC parameters from round $h$ at client $k$.

Local training can leverage diverse quantum-optimized algorithms, such as quantum natural gradient descent (QNGD), which adapts the learning rate based on the geometry of quantum state space using the Fubini–Study metric (Qi, 2022, Qi et al., 2023). The local QNN parameter update with natural gradient is:

$$\theta_{t+1}^{(k)} = \theta_t^{(k)} - \eta \cdot g^{+}(\theta_t^{(k)}) \nabla \mathcal{L}(\theta_t^{(k)})$$

where $g^+$ is the pseudo-inverse of the block-diagonal Fubini–Study metric tensor.

QFL can be realized over both classical and quantum networks. In some architectures, blockchain technology is used for decentralized aggregation and auditability, avoiding single points of failure and strengthening integrity against adversarial attacks (Gurung et al., 2023).

2. Quantum Data Handling and Encoding Techniques

A unique feature of QFL is its ability to process quantum-native data—data produced by quantum sensors or other quantum systems, often encoded as multi-qubit pure states or mixed states. In absence of public federated quantum datasets, custom datasets are generated via local quantum circuit simulations (e.g., cluster states constructed with Cirq and TensorFlow Quantum), with hierarchical formats that partition samples by client (Chehimi et al., 2021).

Classical data (such as images, genomics, or sensor readings) require specialized encoding to map features into quantum states that can be processed by QNNs. Encoding approaches include:

• Amplitude encoding: Encodes an entire data vector into the amplitudes of a quantum state, using an exponentially compact number of qubits. Efficient for high-dimensional data but sensitive to noise.
• Angle encoding: Each classical feature value parametrizes a rotation gate on a corresponding qubit; requires a linear number of qubits.
• Basis encoding: Classical bits are mapped directly to computational basis states; less efficient for high-dimensional data.

Selection of encoding impacts model size, training dynamics, and accuracy. For instance, amplitude encoding was used to encode 200-feature genomics data in only 8 qubits for quantum federated cloud experiments (Pokhrel et al., 1 May 2024).

3. Model Architectures and Fusion Strategies

QFL supports diverse quantum model architectures. Early frameworks studied quantum convolutional neural networks (QCNNs), which consist of alternating "quantum convolutional" layers (quasi-local unitary gates), quantum pooling (measurement-driven state reduction), and fully connected layers. More recent work extends QFL to depths of expressivities and modalities unaddressed by earlier methods.

• Entangled Slimmable QNNs (eSQNNs): Multi-depth architectures where each sub-model (of varying depth/complexity) can be individually extracted and aggregated according to channel quality, with entanglement controlled via universal gates. A fidelity-inspired regularizer (‘inplace fidelity distillation’) mitigates inter-depth interference (Yun et al., 2022).
• Multimodal Fusion with Entanglement: Newer QFL frameworks process multiple data modalities per client (e.g., audio, image, text), with each modality stream handled by a dedicated PQC and fused via an intermediate quantum fusion layer. Trainable entangling gates couple the streams, yielding a fused state:

$$|\Psi_{\text{out}}^{k}(\boldsymbol{\theta}^k)\rangle = U_{\text{fusion}}^k(\boldsymbol{\theta}_{\text{fusion}}^k) \left( \bigotimes_{m=1}^M |\psi_{m,\text{out}}^k(\boldsymbol{\theta}_m^k)\rangle \right)$$

This fusion leverages quantum entanglement to model cross-modal correlations at the quantum state level, improving performance on real-world multimodal tasks (Pokharel et al., 10 Jul 2025).

• Temporal Quantum Models: Federated QLSTM (quantum long short-term memory) models allow distributed approximation of temporal functions in quantum sensor networks, composing LSTM-like architectures from quantum circuit submodules (Chehimi et al., 2023).

4. Communication, Privacy, and Security Considerations

A central challenge in QFL is maintaining data privacy and minimizing communication costs. QFL inherently avoids transmitting raw data, mitigating privacy risks. Practical frameworks have introduced further safeguards:

• Encrypted QFL: The CryptoQFL approach employs a two-stage encryption pipeline, combining quantum one-time-pad encryption of QNN gradients with quantum homomorphic encryption of encryption keys. Ternary quantization of gradients and optimized quantum adder circuits minimize bandwidth and aggregation latency (Chu et al., 2023).
• Quantum Key Distribution (QKD) and Secure Multiparty Protocols: Integration of quantum-secure communication, such as QKD, provides information-theoretic security superior to classical encryption (Mathur et al., 9 Apr 2025, Chehimi et al., 2023).
• Blockchain Integration: Peer-to-peer and smart-contract-based blockchains eliminate central points of failure, ensure transparent model aggregation, and underpin applications such as decentralized finance in virtual environments (Gurung et al., 2023).

QFL models have also been extended to operate with classical clients that lack quantum hardware. The CC-QFL framework employs shadow tomography on the server to construct classical shadows of quantum states, enabling classical clients to calculate local gradients with respect to observables, thereby including them in quantum model training (Song et al., 2023).

5. Robustness to Noise, Channel Variability, and Adversarial Threats

Quantum noise—stemming from decoherence, faulty gates, and measurement errors—presents major obstacles to effective QFL. Noise heterogeneity across devices disrupts model convergence and can dominate gradient signals.

• Sporadic Federated Learning (SpoQFL): SpoQFL dynamically weights or skips local model updates based on estimated noise-induced gradient deviations:

$x_{n,k}^t = \exp(-\gamma |\xi_{n,k}^t|)$

Updates with significant noise (low $x_{n,k}^t$) are suppressed or skipped, thereby avoiding propagation of instability and enhancing training stability and accuracy (Rahman et al., 15 Jul 2025).

• Adaptive Aggregation and Channel Sensitivity: Frameworks such as SlimQFL and eSQFL dynamically send either partial or full parameter sets based on channel conditions or adapt fusion depth based on received sub-models (Yun et al., 2022, Yun et al., 2022).
• Adversarial Robustness: Quantum Federated Adversarial Learning (QFAL) incorporates local adversarial training with federated averaging, shielding global models against adversarial perturbations. Empirically, partial adversarial coverage (20–50% of clients) improves robustness, but a trade-off with clean accuracy persists (Maouaki et al., 28 Feb 2025).

6. Evaluation, Benchmarks, and Practical Validation

QFL frameworks have been validated via simulations and, increasingly, via cloud-based quantum simulators and real quantum hardware (e.g., IBM QPUs).

• Example tasks include binary and multi-way MNIST image classification, Bessel/Struve function approximation for quantum sensing, genomic and sentiment classification, and emotion detection in multimodal datasets (Chehimi et al., 2021, Innan et al., 16 Mar 2024, Chehimi et al., 2023, Pokharel et al., 10 Jul 2025).
• Aggregation schemes explored include simple averaging, weighted averaging, accuracy-based selection (“best-pick”), and knowledge distillation (as in LLM-QFL with LLMs) (Pokhrel et al., 1 May 2024, Gurung et al., 24 May 2025).
• Performance metrics encompass accuracy, convergence rate (number of communication rounds), loss, robustness against noise and adversarial attack, and, in decentralization contexts, blockchain throughput and auditability.

Notably, the use of federated quantum natural gradient descent (FQNGD) accelerates convergence and reduces the number of required communication rounds compared to classical optimizers, achieving accuracy advantages above standard methods in experiments (Qi, 2022, Qi et al., 2023).

7. Open Research Directions and Future Perspectives

Major directions for future research include:

• Scalability to Large and Heterogeneous Networks: As quantum devices scale, new aggregation, error correction, and adaptive optimization strategies are needed to support increased circuit depth and qubit counts while mitigating noise and hardware heterogeneity (Chehimi et al., 2023, Mathur et al., 9 Apr 2025).
• Hybrid Quantum–Classical Integration: Partitioning computation to maximize quantum speedup while leveraging mature classical infrastructure, including model-driven engineering and abstraction tools, is recognized as essential for broader adoption (Moin et al., 2023, Mathur et al., 9 Apr 2025).
• Advanced Security and Privacy Protocols: Methods such as gradient hiding with entanglement, differential privacy with quantum noise, blind quantum computing, and cryptographically secure multiparty protocols remain active research areas (Mathur et al., 9 Apr 2025).
• Multimodal and Real-world Deployment: The extension of QFL frameworks to handle multimodal data, missing modalities (via "Missing Modality Agnostic" MMA mechanisms), and applications in domains such as quantum-enabled IoT, healthcare, finance, and the metaverse is ongoing (Pokharel et al., 10 Jul 2025, Gurung et al., 2023).
• Benchmarking and Standardization: As the field matures, there is a strong demand for standardized datasets, performance benchmarks, explainability, and comprehensive evaluation criteria specific to QFL (Ren et al., 2023, Chehimi et al., 2023).
• Integration of Foundation Models: LLM-QFL demonstrates that LLMs, adaptively fine-tuned within QFL, can act as centralized or distributed reinforcement agents to optimize communication, convergence, and client selection, signifying a direction where quantum-native and AI-native paradigms converge (Gurung et al., 24 May 2025).

Quantum Federated Learning thus represents a multidisciplinary convergence of quantum information science, distributed machine learning, cryptographic security, and complex systems engineering, with early empirical results indicating both practical viability and a wealth of open theoretical and implementation challenges.