Quantum-Enhanced AI Boosts Accuracy by 10.22 Per Cent with Privacy Safeguards

Researchers are tackling the challenge of maintaining machine learning performance when training models across multiple, separate datasets without compromising data privacy. Yueheng Wang, Xing He, and Zinuo Cai, from Jiangnan University and Shanghai Jiao Tong University, alongside Ruhui Ma and Yuan Liu, present FEDCOMPASS, a novel framework designed to improve hybrid classical-quantum federated learning. Their work addresses performance degradation caused by uneven data distribution across these decentralised systems by employing a layered aggregation approach that groups clients based on data similarity and utilises adaptive optimisation techniques. Demonstrating improvements of up to 10.22% in test accuracy and enhanced convergence stability on benchmark datasets, FEDCOMPASS represents a significant advance in enabling robust and efficient federated learning for increasingly complex models.

This research addresses a critical limitation in current federated learning approaches, namely performance degradation when dealing with non-IID data, where data distributions vary significantly across different clients. FEDCOMPASS introduces a two-tiered aggregation strategy, separating the processing of classical feature extractors from quantum parameters to optimise performance and stability. The core innovation lies in the use of spectral clustering to group clients based on the similarity of their data class distributions, enabling cluster-wise aggregation for the classical components of the model. For the quantum parameters, researchers implemented circular mean aggregation, a technique that accounts for the periodic nature of these values, combined with adaptive optimisation to ensure stable global updates. This approach mitigates the parameter period mismatch that can occur with direct averaging, preventing optimisation conflicts between the classical and quantum modules. Extensive experimentation on three benchmark datasets, MNIST, Fashion-MNIST, and CIFAR-10, demonstrates that FEDCOMPASS improves test accuracy by up to 10.22% compared to six established federated learning baselines. The study highlights a significant advancement in addressing data heterogeneity within federated learning environments. By dynamically clustering clients and employing tailored aggregation methods for both classical and quantum layers, FEDCOMPASS achieves enhanced convergence stability under non-IID conditions. This layered approach ensures that the global model benefits from the collective knowledge of diverse clients without being hampered by biased local models. The framework’s ability to preserve privacy while improving both accuracy and stability positions it as a promising solution for real-world applications in areas such as mobile healthcare, the Internet of Things, and distributed sensing. FEDCOMPASS’s cluster-based aggregation for classical feature extraction utilises client category statistics, performing weighted aggregation within clusters to generate cluster-level feature extractors. This process avoids uploading raw data, maintaining data privacy throughout the learning process. Furthermore, the circular mean aggregation applied to quantum parameters, combined with adaptive optimisation, ensures robust global updates by accounting for the unique characteristics of quantum data. The comprehensive experimental validation confirms FEDCOMPASS’s advantages in accuracy, stability, and privacy preservation within hybrid classical-quantum federated learning settings, paving the way for more efficient and reliable distributed machine learning. Client partitioning via distributional similarity and weighted model aggregation improves personalization and efficiency Spectral clustering underpinned the initial stage of the FEDCOMPASS methodology, grouping clients based on the similarity of their class distribution vectors. Each client first computed a category distribution vector representing the proportion of samples belonging to each class within its local dataset. The server then collected these vectors and constructed a similarity matrix, calculating the Jensen-Shannon divergence between each pair of client vectors to quantify distributional differences. Applying spectral clustering to this matrix partitioned the clients into distinct clusters, facilitating subsequent weighted aggregation within each group. Following client clustering, weighted aggregation generated cluster-level classical feature extractors. Within each cluster, the server averaged the parameters of the classical feature extraction layers from all constituent clients, weighting each client’s contribution according to its data distribution. This process mitigated model deviation arising from non-IID data by ensuring that the aggregated feature extractors were representative of the data characteristics within each cluster. The resulting cluster-level models were then prepared for integration with updated quantum parameters. To address the periodicity inherent in quantum parameters, a circular mean aggregation technique was implemented. This method treated quantum parameters as points on the unit circle, performing angle wrapping to maintain parameter values within the range of negative pi to positive pi. Averaging these angles provided a robust estimate of the global quantum parameters, avoiding the parameter period mismatch that can destabilize training. An adaptive optimizer was then applied to these aggregated quantum parameters, further refining the global updates and promoting stable convergence. Extensive experimentation validated FEDCOMPASS on the MNIST, Fashion-MNIST, and CIFAR-10 datasets.

Results demonstrated improvements in test accuracy, reaching up to 10.22% over six established federated learning baselines, and enhanced convergence stability under non-IID conditions. The framework’s performance consistently highlighted its advantages in accuracy, stability, and privacy preservation within hybrid classical-quantum federated learning environments. FEDCOMPASS enhances federated learning accuracy and stability via spectral clustering and weighted aggregation of model updates Improvements of up to 10.22% in test accuracy were achieved through the FEDCOMPASS framework across three benchmark datasets. This layered aggregation framework addresses performance degradation in hybrid classical-quantum federated learning under non-IID data conditions. The research demonstrates enhanced convergence stability alongside the accuracy gains, consistently surpassing six established federated learning baselines. FEDCOMPASS utilizes spectral clustering to group clients based on similarity in class distribution, enabling cluster-wise aggregation for classical feature extractors. This clustering process relies on a similarity matrix, calculated using both Jensen, Shannon divergence and sample size discrepancy, to effectively partition heterogeneous client groups. The similarity metric considers distribution divergence and sample size differences, with hyperparameters λ1 and λ2 balancing their respective weights in the calculation. For the classical layer, weighted average aggregation is performed within each identified cluster, using local sample sizes as weights to generate cluster-level feature extractors. This approach ensures that the model adapts effectively to data heterogeneity without uploading raw data, preserving privacy. The resulting cluster-shared classical model benefits from the refined aggregation process. Quantum parameter aggregation employs a circular mean method, resolving inconsistencies arising from the periodicity of rotation angles. This circular mean aggregation is applied to each dimension of the quantum parameters, utilizing client weights determined by their respective sample sizes. An adaptive optimization strategy, incorporating FedAdam, further enhances the stability of global convergence, resulting in robust global updates to the quantum classifier. The algorithm computes client weights ωi as the ratio of local sample size ni to the total sample size across all clients.

Mitigating Data Heterogeneity in Hybrid Classical-Quantum Federated Learning requires careful consideration of system constraints FEDCOMPASS, a layered aggregation framework, addresses performance degradation in hybrid classical-quantum federated learning caused by non-IID data distributions. The framework utilises spectral clustering to group clients based on similarities in their data distributions, enabling cluster-specific aggregation of classical feature extractors. Furthermore, it employs circular mean aggregation, alongside adaptive optimisation, to stabilise global updates of quantum parameters, which are inherently periodic in nature. Experimental results, conducted on three benchmark datasets, demonstrate that FEDCOMPASS improves test accuracy by up to 10.22% and enhances convergence stability when dealing with non-IID data. This performance surpasses that of six established federated learning baselines, indicating a substantial advancement in handling data heterogeneity within hybrid learning systems. The authors acknowledge that the method’s effectiveness relies on appropriate selection of the clustering parameters and the adaptive optimisation rate. Future research could explore automated parameter tuning and investigate the framework’s scalability to even more complex and diverse federated learning scenarios, potentially including asynchronous client updates or more sophisticated quantum parameter encoding schemes. 👉 More information 🗞 Fedcompass: Federated Clustered and Periodic Aggregation Framework for Hybrid Classical-Quantum Models 🧠 ArXiv: https://arxiv.org/abs/2602.03052 Tags: