Towards Personalized Quantum Federated Learning for Anomaly Detection

TL;DR

The paper tackles anomaly detection in privacy-sensitive, distributed settings where anomalies are context-dependent and labeled data are scarce. It introduces personalized quantum federated learning (PQFL), which couples a variational quantum eigensolver (VQE) based quantum model with a regularization-based personalization layer to accommodate heterogeneous encoding schemes and data distributions across $N$ clients over $K$ global rounds and $T$ local epochs, including a proximal term $\lambda (\boldsymbol{w}_{n,k} - \boldsymbol{w}_{global})$ to balance local adaptation with global coherence. Key contributions include the PQFL architecture, a proximal personalization SGD update, a VQE-based QFL workflow with amplitude encoding across $D_c$ qubits and $l$ layers, and a complexity analysis showing $O(K N T \cdot GradEval)$ time and $O(N p)$ space. Through extensive simulations on CIFAR-10/100, SVHN, and ImageNet subsets under non-IID distributions, PQFL achieves notable improvements in false/e missing error rates and classification metrics (e.g., AUROC and AUPR), demonstrating faster convergence and robustness to quantum noise. The work highlights the potential of privacy-preserving, heterogeneity-aware quantum learning for scalable anomaly detection in realistic quantum networks, with practical impact on industrial monitoring, healthcare, and security deployments.

Abstract

Anomaly detection has a significant impact on applications such as video surveillance, medical diagnostics, and industrial monitoring, where anomalies frequently depend on context and anomaly-labeled data are limited. Quantum federated learning (QFL) overcomes these concerns by distributing model training among several quantum clients, consequently eliminating the requirement for centralized quantum storage and processing. However, in real-life quantum networks, clients frequently differ in terms of hardware capabilities, circuit designs, noise levels, and how classical data is encoded or preprocessed into quantum states. These differences create inherent heterogeneity across clients - not just in their data distributions, but also in their quantum processing behaviors. As a result, training a single global model becomes ineffective, especially when clients handle imbalanced or non-identically distributed (non-IID) data. To address this, we propose a new framework called personalized quantum federated learning (PQFL) for anomaly detection. PQFL enhances local model training at quantum clients using parameterized quantum circuits and classical optimizers, while introducing a quantum-centric personalization strategy that adapts each client's model to its own hardware characteristics and data representation. Extensive experiments show that PQFL significantly improves anomaly detection accuracy under diverse and realistic conditions. Compared to state-of-the-art methods, PQFL reduces false errors by up to 23%, and achieves gains of 24.2% in AUROC and 20.5% in AUPR, highlighting its effectiveness and scalability in practical quantum federated settings.

Figures (5)

• Figure 1: Overview of the PQFL architecture with $N$ quantum clients and a classical server within the QFL framework. Each client encodes conventional input into quantum states and is trained using local QML models with PQFL SGD update. The server aggregates the parameters of the local models and updates the global model parameters for the next round.
• Figure 2: An illustration of false and missing detection concerns in QFL anomaly detection. In this scenario, photos of cats and dogs are considered normal, but those of other groups are considered abnormal.
• Figure 3: Non-IID data distribution with three types of distribution. (a) is 'Step' data distribution and (b) and (c) are both 'Dirichlet' data distribution with learning rates of 0.1 and 0.01, respectively, across 10 clients. The figure demonstrates that the step is more skewed than Dirichlet and among Dirichlet, the parameter of $\alpha = 0.01$ is more skewed.
• Figure 4: Performance comparison between QFL, QNN, and FL approaches. we compared both FE (%) (a) and ME (%) (b) in anomaly detection where we use CIFAR-10 as normal dataset and CIFAR-100 as anomaly dataset.
• Figure 5: Comparison between different approaches for Anomaly detection, where CIFAR-10 data is used as normal data and CIFAR-100 as anomaly data with 10 clients Non-IID data distribution over 100 global rounds. We compared using false error (FE) in (a), missing error (ME) in (b), AUROC score in (c), and AUPR score in (d). Note that lower FE and ME values and higher AUROC and AUPR scores indicate better performance.