Page 93 - AIH-2-4

Artificial Intelligence in Health Federated learning health stack against pandemics

90% attack accuracy in FL environments. To mitigate supporting equitable participation. Nonetheless, system
these attacks, a membership inference defense mechanism heterogeneity continues to extend training time due to
named DefMIA was introduced, which adds adversarial hardware capabilities and network disparities, increase
perturbations to global model parameters, reducing attack client dropout rates, and introduce latency, all of which
accuracy to approximately 50% without affecting model hinder overall FL performance. These challenges warrant
performance. Overall, GAN-based FL remains susceptible further research.
to accuracy degradation and privacy concerns when
applied to non-IID medical data, necessitating critical 5. Conclusion
evaluation and ongoing refinement. In this article, a hierarchical FL framework is

Model heterogeneity refers to differences across model proposed, designed to enhance scalability, security, and
architectures, capacities, and training approaches across computational efficiency through multilevel aggregation
medical centers. Before FL implementation, architectural and HE, enabling effective responses to future pandemics.
compatibility must be ensured to facilitate effective FL is a decentralized ML paradigm in which each client
knowledge transfer across diverse model structures without trains a model on local data and shares only model updates
compromising performance. Scale adaptability should also with a central server, thus preserving data privacy. The
be enabled in the platform to accommodate the varying hierarchical structure, consisting of client nodes, local
computational capabilities of institutions by enabling servers, and a central server, supports distributed learning
flexible model sizes. In addition, knowledge alignment and efficient communication. Local servers perform
mechanisms are needed to coordinate learning across intermediate aggregation, reducing communication costs
heterogeneous models. Personalization support should and enhancing scalability. The proposed framework
be enabled to address institution-specific requirements or supports dynamic data and can be extended to train
patient groups without compromising the global model’s specialized models at the central unit based on data
integrity or performance. Therefore, to address model characteristics. It ensures secure model updates through
variation, a hierarchical self-distillation approach, such as parameter encryption and seamless integration with
HierarchyFL, is recommended. This method facilitates ByITFL, which incorporates Byzantine fault tolerance and
70
architecture compatibility by enabling resource-limited information-theoretic privacy protection, strengthening
centers to efficiently learn from advanced models deployed the integrity and confidentiality of updates during
at major sites. However, conventional aggregation aggregation.
techniques, such as FedAvg, cannot be directly applied in To address the demands of medical image-based
this setting. In addition, model architecture inconsistencies diagnostics, the framework is adaptable to advanced
reduce knowledge transfer, as some clients may be unable CNN architectures, such as ResNet34, DenseNet121, and
to contribute weights, leading to performance degradation EfficientNet, which have demonstrated high accuracy
and inconsistency. in tasks such as COVID-19 detection. Similarly, in lung
System heterogeneity refers to differences in nodule detection, the incorporation of modular clustering
computational resources, storage capacity, and and transfer learning enhances imaging operations by
communication bandwidth. Strategies to manage these efficiently identifying ROI, reducing resource usage
differences are essential. Resource-conscious participation while improving diagnostic accuracy. On the other
should be prioritized, with computational overhead hand, blockchain-based smart contracts facilitate trusted
adapted to each institution’s hardware and network coordination among participants, while GANs contribute
capacity. Efficiency in communication is also required for to standardized data processing. Experiments conducted
optimal data transfer, particularly within low or unstable using the Flower framework and TenSEAL on the MNIST
bandwidth settings. Furthermore, storage efficiency should dataset demonstrated improved model accuracy, reduced
accommodate institutions with limited infrastructure training time per round, and lower communication
to ensure broad participation. Power efficiency is also overhead due to the scope of optimized intermediate
important, especially for edge devices, to reduce power aggregation.
consumption with consistent performance within clinical Overall, the proposed framework offers a scalable,
settings. A framework incorporating client selection privacy-preserving, and computation-efficient solution
through reinforcement learning and resource-aware for digital healthcare. These advancements collectively
metrics, such as that proposed in personalized FL, is position the proposed framework as a robust approach for
71
recommended. This framework performs federation tasks addressing future healthcare challenges and responding
dynamically according to each center’s available resources, effectively to potential pandemics.

Volume 2 Issue 4 (2025) 87 doi: 10.36922/AIH025080013