Page 88 - AIH-2-4

Artificial Intelligence in Health Federated learning health stack against pandemics

scoring evaluates client updates with a polynomial ReLU 2.7. Deep-learning protocol
function: In the previous section, the FL architecture was outlined.

gg , However, specific deep-learning models that might be
TS ReLU 0 i (XIII) integrated to address future pandemics were not discussed.
i
 g  0 . g  i As future pandemics are uncertain, to illustrate the point,
COVID-19 and lung cancer were taken as benchmarks for
Where TS is the trust score for client i, g is the reference potential adaptations of deep-learning protocols.
i
0
gradient, and g is the encoded gradient. Here, the dot
22
i
29
product gg, i gives the scalar projection, measuring In this context, the study by Gogineni et al. was chosen
0
how directionally aligned the two gradient vectors are, as a reference. The study investigated the potential of deep-
learning models for automated COVID-19 detection using
while the Euclidian norm in the denominator, which chest X-ray images, presenting a promising alternative
measures the magnitude of the vectors g and g , to the current gold standard – reverse-transcription
0
i
normalizes this measure so that it lies in (−1, 1). The polymerase chain reaction (RT-PCR) test. This choice has
updates with TS < 0.1 can be discarded. For LCC-based two distinct merits. First, COVID-19 represents the most
i
aggregation, global gradients are computed as:
recent pandemic for benchmarking. Second, choosing
N TS g. 2 images as input data type. This choice is crucial as image
,.
(~
g global i1 N i i N 00 1 g global (XIV) data is representative of various real-world medical datasets
i1 TS i and therefore can be used as a reliable proxy (since the
where η is the re-randomization noise drawn from exact nature of future pandemics is unknown). Moreover,
a zero-mean Gaussian distribution  with a standard images are one of the most complex and prevalent data
deviation equal to 10% of the global gradient’s ℓ₂ norm. types in the medical arena; therefore, demonstrating with
Such noise interferes with deterministic patterns across proxy images offers one of the most effective benchmarking
training iterations, making gradient memorization or strategies to mimic real-world complexities. Although
reverse engineering less likely without compromising videos represent more complex data types, they are far less
frequent in the medical context and, in a crude sense, can
update fidelity. 14,27
be considered as a sequential stacking of images with time-
2.6. Integration of FedML–HE with ByITFL for series data (audios, if any) added in an additional channel.
enhanced privacy In the study by Gogineni et al., several CNN
29
30
The combination of FedML–HE’s selective HE and architectures were implemented, including ResNet34,
23
33
34
ByITFL’s Byzantine-resilient architecture offers an SeResNext50, 31,32 DenseNet121, and EfficientNet.
28
optimally balanced solution to FL’s dual challenges of These models were chosen for their distinct advantages
privacy and security, providing a secure and reliable in image classification tasks. ResNet34 utilizes skip
network architecture. By using FedML–HE’s parameter- connections, allowing for efficient training of deep
level encryption on sensitive gradients that are detected networks, while SeResNext50 incorporates squeeze-
through ByITFL’s trust scoring, the hybrid solution is and-excitation blocks, which recalibrate channel-wise
28
able to effectively protect important model updates against feature responses for improved representational capacity.
23
inversion attacks while maintaining Byzantine resilience. Meanwhile, DenseNet121, with its dense connections
Empirical evaluations have demonstrated that selective between layers, facilitates feature reuse and enhances
encryption can cut down on communication overhead by information flow. Finally, EfficientNet models are
up to 10 times for large models such as ResNet-50 without designed using neural architecture search, optimizing the
23
34
a loss in malicious client detection accuracy. Earlier work balance between accuracy and computational efficiency.
35
on Byzantine-resilient secure aggregation architectures Transfer learning, using the ImageNet dataset, was
points to the possibility of combining cryptographic employed to improve model performance on the relatively
36
privacy with adversarial robustness. In addition, FedML- limited medical image dataset. A learning rate scheduler
HE’s performance-optimized encryption pipeline is and a one-cycle training policy were also implemented
optimally compatible with ByITFL’s computationally for better convergence and generalization.
limited architecture. The hybrid solution helps prevent The models’ performance demonstrated encouraging
privacy leakage risks in trust-based aggregation while results. ResNet34 and DenseNet121 achieved the
29
maintaining the system’s ability to filter poisoned updates, highest overall accuracy of 94.09% in classifying images
a circumstance verified by cross-institutional medical FL as COVID-19, normal, or pneumonia. This accuracy is
trials. 23 considerably higher than the typical 70 – 80% sensitivity
Volume 2 Issue 4 (2025) 82 doi: 10.36922/AIH025080013