🤖 machine learning

A Drift Stable Quantum Federated Learning for Intelligent Services

This paper proposes DUQFL-Prox, a drift-stable quantum federated learning framework that combines adaptive unfolded SPSA updates with a proximal term and a lightweight controller to enhance stability, generalization, and client fairness in heterogeneous distributed intelligent services.

Original authors: Shanika Iroshi Nanayakkara, Shiva Raj Pokhrel

Published 2026-07-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shanika Iroshi Nanayakkara, Shiva Raj Pokhrel

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where your phone, your smartwatch, and your neighbor's laptop all want to learn together to solve a big problem, like spotting a bank fraud or identifying a rare genetic trait. But there's a catch: none of them want to share their private data. They are like a group of detectives who can't show each other their secret case files. This is the world of Federated Learning, a clever way for computers to learn as a team without ever swapping their private notes. Now, imagine taking this team effort and giving them superpowers from the quantum world—the realm of atoms and subatomic particles where things can be in many places at once. This is Quantum Federated Learning. It promises to solve complex problems faster and more securely, but it's a bit like trying to conduct an orchestra where every musician is playing on a shaky, noisy instrument that sometimes forgets the notes. The big question scientists are asking is: How do we get these quantum computers to work together smoothly without one loud musician ruining the song for everyone else?

This paper introduces a new team captain for this quantum orchestra called DUQFL-Prox. The authors suggest that the current way of training these quantum models is too rigid, like forcing every musician to play the exact same notes at the exact same speed, regardless of how their instrument is behaving. This often leads to "client drift," where some musicians get so carried away with their own practice that they drift off-key, making the final group performance messy and unfair. To fix this, the researchers proposed a system where each quantum computer (or "client") doesn't just follow a fixed script. Instead, they use a smart, adaptive coach that watches the musician's progress in real-time and tweaks the instructions on the fly.

Think of it like a video game where the difficulty adjusts automatically. If a player is struggling, the game slows down; if they are breezing through, it speeds up. In this system, the "coach" (a lightweight controller) constantly checks how well the local quantum model is doing and adjusts the learning steps. But there's a safety net, too: a "proximal" rule that gently pulls the local model back toward the group's main goal if it starts to wander too far. It's like a bungee cord that lets you explore but keeps you from falling off the cliff. The researchers also added a "best-of" rule: instead of just sending the very last version of their model to the group, each client sends back the version that performed best during their practice session, skipping the ones that got too messy at the end.

When the team tested this new method on two very different challenges—spotting bank fraud and classifying DNA sequences—they found some promising signs. In the bank fraud tests, which were tricky because the "bad" cases were very rare, DUQFL-Prox didn't just get a high overall score; it made sure that every single client (every bank) performed well, not just the lucky ones. It reduced the gap between how well the model did in practice versus how well it did in the real world, suggesting it was less likely to be fooled by fake patterns. In the DNA tests, while another method got a slightly higher top score, DUQFL-Prox was the most consistent and fair, ensuring that no single client was left behind. The authors ran these experiments on computer simulations of quantum computers and even checked if the final models could run on real, noisy quantum hardware from IBM. The results suggest that this adaptive, drift-stable approach helps quantum teams learn more reliably and fairly, even when their data is messy and their instruments are a bit shaky. It's not a magic bullet that solves everything instantly, but it suggests a path toward making these futuristic quantum services more trustworthy for everyone.

Technical Summary: DUQFL-Prox for Drift-Stable Quantum Federated Learning

Problem Statement
Quantum Federated Learning (QFL) enables distributed clients to train Quantum Neural Networks (QNNs) without centralizing raw data, a capability critical for privacy-sensitive intelligent services such as financial fraud detection and genomic classification. However, existing QFL pipelines face significant challenges in heterogeneous environments. Local optimization of QNNs is inherently noisy, non-convex, and sensitive to optimizer configurations. When client data is non-independent and identically distributed (non-IID), fixed local optimization schedules (e.g., standard SPSA with static hyperparameters) often lead to unstable local updates, excessive client drift, and unfair performance across clients. Furthermore, strategies that exclude "straggler" clients to improve efficiency can introduce representativeness bias, particularly detrimental in domains where rare or domain-specific data distributions are crucial. The core problem is to stabilize local QNN optimization trajectories in heterogeneous settings to ensure aggregation compatibility, fairness, and reliable generalization without discarding difficult clients.

Methodology: DUQFL-Prox
The paper proposes DUQFL-Prox (Deep-Unfolded Quantum Federated Learning with Proximal Regularization), a framework designed to stabilize QFL through a bilevel learning structure:

Deep-Unfolded Local Optimization: Instead of treating local training as a fixed routine, the method decomposes local QNN training into $K$ unfolded optimization blocks. At each step $k$ , a lightweight, shared meta-controller generates adaptive SPSA (Simultaneous Perturbation Stochastic Approximation) hyperparameters—specifically the learning rate ( $\eta$ ) and perturbation scale ( $\delta$ ). These are derived from an optimization-state feature vector ( $z$ ) that includes unfold progress, loss behavior, parameter displacement, and client context.
Proximal Drift Control: To counteract client drift under non-IID data, a proximal regularization term is added to the local objective function. This term penalizes excessive deviation from the broadcast global model ( $\theta^{(t)}$ ), ensuring that while clients adapt locally, their updates remain compatible with global aggregation. The local objective becomes $L_i^{prox}(\theta) = L_i(\theta) + \frac{\mu}{2}\|\theta - \theta^{(t)}\|_2^2$ .
Validation-Based Best-Unfold Selection: Clients do not necessarily upload the final state of their local training. Instead, they evaluate the validation loss across all unfolded states and upload the checkpoint ( $\theta_{i, \star}$ ) that minimizes validation loss. This prevents over-aggressive later updates from degrading the global model.
Outer Meta-Objective and Controller Adaptation: The shared controller is not static; it is periodically refined using an outer meta-objective evaluated on post-aggregation global validation performance. This meta-loss includes terms for global validation loss, client-level fairness, communication cost, and optimization stability. The controller parameters are updated using an outer-loop SPSA step, effectively learning "how to optimize" the local QNNs to maximize global and fair performance.

Key Contributions

Adaptive Framework: The introduction of DUQFL-Prox, which models local QNN training as an adaptive, multi-step deep-unfolded process rather than a fixed optimizer routine.
Controller-Driven SPSA: A mechanism where a shared controller dynamically generates step-specific SPSA learning rates and perturbation scales based on local optimization states, enabling adaptation across clients and communication rounds.
Drift and Fairness Control: The integration of proximal regularization and validation-based best-unfold selection to reduce client drift, prevent over-specialization, and improve aggregation compatibility.
Bilevel Learning Structure: An outer-loop adaptation mechanism that aligns local optimization behavior with post-aggregation global performance, learning to optimize the optimization process itself.

Experimental Results
The framework was evaluated on two intelligent service tasks: Bank Account Fraud (BAF) detection (highly imbalanced, non-IID) and Genomic classification.

BAF Dataset: DUQFL-Prox achieved the highest final global accuracy (~~0.65) and mean client test accuracy (~~0.64), significantly outperforming FedProx-QFL and Default-QFL. Crucially, it reduced the train-test gap to near-zero and achieved the lowest client fairness gap (~0.019), indicating balanced performance across clients. It also demonstrated more stable late-epoch behavior in ROC-AUC, PR-AUC, and MCC compared to baselines.
Genome Dataset: While FedProx-QFL achieved slightly higher final global accuracy (~~0.85), DUQFL-Prox achieved the highest mean client test accuracy (~~0.80), the lowest train-test gap, and the lowest fairness gap. This highlights its superior generalization and stability across heterogeneous clients.
Hardware Validation: Selected global checkpoints trained via DUQFL-Prox were executed on real IBM Quantum hardware. The results showed that DUQFL-Prox checkpoints maintained a more stable trajectory on real hardware compared to the Default-QFL baseline, retaining alignment with simulator results despite device noise.

Significance and Claims
The paper claims that DUQFL-Prox offers a robust solution for intelligent services requiring privacy, fairness, and reliability in heterogeneous distributed environments. By stabilizing local optimization trajectories through deep unfolding and proximal control, the framework allows difficult or resource-limited clients to contribute useful updates without being excluded. The authors position the work not as a full convergence guarantee for arbitrary non-convex QFL systems, but as a characterization of stability, drift-control, and meta-optimization behaviors that support more reliable and fair intelligent services. The results suggest that deep-unfolded quantum federated learning can effectively manage the trade-offs between local adaptability and global consistency in noisy quantum settings.

Technical Summary: DUQFL-Prox for Drift-Stable Quantum Federated Learning

More like this