Page 83 - AIH-2-4

Artificial Intelligence in Health Federated learning health stack against pandemics

privacy-preserving, hierarchically structured FL approach 1.2. Technical preliminaries
to address these challenges. In this section, we discuss some technical preliminaries used
We present a secure and scalable FL framework in the architectural design of the proposed hierarchical FL,
using a hierarchical structure and homomorphic such as static and dynamic datasets, hierarchical learning,
16
encryption (HE) for privacy-preserving, collaborative and HE. A static dataset does not change over time,
17
model training. Then, we discuss the main features of whereas a dynamic dataset is continuously or periodically
14
Byzantine-resilient information-theoretic FL (ByITFL), updated with new data. HE is a privacy-preserving
which offers privacy-preserving aggregation and robust technique that enables computations, such as addition and
protection against communication-based attacks in FL multiplication, to be performed directly on encrypted data.
through standard security techniques. Subsequently, we When data or model parameters are transmitted to a server,
illustrate an implementation of a deep-learning protocol there is a risk of interception or tampering by attackers,
based on modern convolutional neural network (CNN) potentially compromising model integrity and predictions.
architectures that efficiently detect past pandemics like In the proposed framework, we employ HE to encrypt
COVID-19 and diseases like lung cancer, with the aim model parameters before transmission. These encrypted
of improving accuracy and computational efficiency by parameters are then aggregated on the server using
leveraging transfer learning, clustering algorithms, and homomorphic addition and multiplication operations,
region-of-interest (ROI) analysis. Blockchain/smart ensuring both data privacy and secure computation. FL
contracts play a key role in establishing digital trust and frameworks may suffer from scalability limitations due
maintaining transparency, both of which are critical for to communication bottlenecks that arise when a large
success in a global endeavor, such as the fight against number of clients frequently transmit model updates to a
pandemics. Unfortunately, they are not free from external centralized server. This issue becomes more pronounced in
vulnerabilities. We discuss detecting vulnerabilities in smart distributed environments, such as health care.
contracts using an improved averaged stochastic gradient To address this, we employ hierarchical learning.
18
descent weight-dropped long short-term memory (AWD- Hierarchical learning is a structured learning process
LSTM) model, which achieves high accuracy and F1 score organized across multiple levels, where higher levels
by incorporating opcode review analysis and addressing aggregate and refine knowledge from lower levels to
class imbalance in smart contract datasets. In Section 3, improve performance and scalability. It has a multi-tiered
we experimentally evaluate our proposed framework structure comprising a central server along with multiple
and present a detailed cost analysis of the hierarchical local servers at each level. This structure efficiently
FL framework, showing its scalability through efficient handles gradient aggregation and transmission in
15
communication and computation at client, local server, and parallel, which reduces overall communication overhead
central server levels. Communication costs are optimized and ensures privacy. Figure 1 represents a hierarchical
using hierarchical data aggregation, while computation learning architecture suitable for deployment in the global
costs are minimized through localized operations, ensuring healthcare system.
scalability and security. In Section 4, we address the issue
of non-independent and identically distributed (non-IID) We experimentally evaluated our proposed framework
datasets and heterogeneity considerations, both of which using a proxy image classification dataset, implementing
a hierarchical FL architecture and transmitting model
are critical for the efficient implementation of our idea.
parameters through HE. We compared its performance
The FL model that we have described (including against standard FL lacking hierarchical structure and
hierarchical FL, ByITFL, and HE) is more realistic, as it encrypted communication. Our experimental results
takes into consideration real-world applications. It is also demonstrate that the hierarchical framework achieved
suggested to use institutional incremental learning and higher model accuracy and reduced training time compared
10
cyclic institutional incremental learning. FL achieves to generic FL. The hierarchical structure effectively
10
a better rate of model improvement than data-private distributes the computational and communication load,
collaborative learning methods. Moreover, to compare the enabling faster convergence. Additionally, the use of HE
rates of model improvement, a global validation “dice-over- ensures that sensitive information remains confidential
epoch” (where the model trains over epochs and the metric throughout the training process, thereby satisfying
is computed on both training and validation data) for all privacy requirements without significantly compromising
collaborative methods showed that FL training converges efficiency. The proposed model is capable of incorporating
relatively quickly to the same performance as collaborative both static and dynamic data. However, future pandemics
data sharing training. may involve highly distributed or heterogeneous data types.

Volume 2 Issue 4 (2025) 77 doi: 10.36922/AIH025080013