Adaptive Aggregation with Two Gains in Quantum Federated Learning Enables Robust Performance in Heterogeneous Networks

Federated learning, a promising technique for collaborative machine learning, frequently suffers performance losses when deployed across real-world networks with varying device capabilities and unreliable connections. Shanika Iroshi Nanayakkara and Shiva Raj Pokhrel, from Deakin University, address this challenge with a new approach called A2G-QFL, or Adaptive Aggregation with Two Gains in Quantum Federated learning. Their research introduces a framework that dynamically adjusts how individual contributions are combined, taking into account both the geometric relationship between models and the quality of service provided by each participating device. By jointly regulating these factors, A2G-QFL demonstrably improves stability and accuracy in challenging, heterogeneous network conditions, and importantly, encompasses existing methods like FedAvg as specific instances of its broader approach.

Geometry and Quality Gain for Aggregation

The research team developed a novel aggregation framework, A2G, to enhance federated learning across diverse and potentially unstable communication networks, including those integrating quantum and classical technologies. This method tackles performance degradation caused by variations in client quality, communication reliability, and geometric differences between local and global models. A2G operates by jointly regulating geometric blending through a ‘geometry gain’ and modulating client importance using a ‘QoS gain’ derived from measures of teleportation fidelity, latency, and instability. Clients receive a global model parameter and execute a local optimizer to generate a local model and, optionally, a stochastic gradient estimate, simultaneously measuring key physical-layer quality of service indicators: teleportation fidelity, communication latency, and channel instability variance.

These measurements, along with the local model update and dataset size, are transmitted to the server. Researchers implemented a procedure to ensure consistent data collection across the network and computed QoS-sensitive trust coefficients, weighting each client’s contribution to the aggregation process. This involved calculating a QoS factor based on fidelity, latency, and instability, regulated by adjustable parameters, then normalizing this factor by the data-size proportion of each client to create a trust score. This score is further normalized across all clients, resulting in weights that prioritize reliable, high-performing nodes, allowing for tunable control over the sensitivity to each QoS indicator. The team implemented geometry-controlled aggregation, combining the QoS-weighted gradient terms with a geometry gain, effectively interpolating between standard Euclidean aggregation and a more sophisticated geometry-aware averaging on the underlying parameter manifold. By adjusting the geometry gain, the researchers could adapt the aggregation process to the specific characteristics of the model and communication network, recovering existing methods like FedAvg and QoS-aware averaging as special cases.

Adaptive Federated Learning with Quantum Links

The research presents A2G, an adaptive aggregation framework designed to improve federated learning systems operating across networks with varying client quality and potential quantum communication links. This work addresses performance degradation caused by factors like unreliable connections and differences in client capabilities. A2G uniquely combines quality of service (QoS) gains, derived from metrics like teleportation fidelity, latency, and instability, with geometry gains that regulate how local and global models are blended. Experiments on a hybrid testbed and datasets including Breast-Lesions-USG demonstrate that A2G significantly enhances both stability and accuracy.

With a geometry gain of 0. 05, the system achieved a best accuracy of 68. 25 percent and a final accuracy of 63. 49 percent on the Breast-Lesions-USG dataset, consistently outperforming the standard FedAvg baseline. Moderate geometry gains consistently deliver superior results; a gain of 0.

10 yielded a best accuracy of 65. 08 percent, while higher values showed systematic degradation in performance. Further investigation into noisy communication links, modeled as a bit-flip channel, revealed A2G’s robustness. Under medium noise conditions, A2G maintained strong performance, and even with increased noise, it still outperformed the FedAvg baseline. The team addressed challenges arising from varying client quality and communication constraints by developing a method that jointly regulates geometric blending and client importance through the implementation of geometry gains and quality of service (QoS) gains, derived from metrics such as teleportation fidelity, latency, and instability. The A2G update rule demonstrably improves stability and accuracy in challenging conditions, while also generalizing existing methods like FedAvg, QoS-aware averaging, and manifold-based aggregation. Experiments conducted on a hybrid testbed confirm the framework’s effectiveness in mitigating both client drift and bias induced by quality of service variations. The modular structure of A2G and its compatibility with various optimizers and model classes position it as a promising foundation for advanced distributed and quantum-enabled learning frameworks, though future work will focus on establishing convergence guarantees and addressing assumptions of smoothness and bounded variance.