A Federated Learning System with Biometric Medical Image

Authentication for Alzheimer's Diagnosis

Francesco Castro

, Donato Impedovo

and Giuseppe Pirlo

Univeristy of Bari Aldo Moro, Via E. Orabona, 4, Bari, 70125, Italy

Keywords: Federated Learning, Alzheimer’s Disease, Medical Image, Data Poisoning, Secure Healthcare System.

Abstract: There are concerns within the medical/scientific community about the use of machine learning models for

disease diagnosis from medical images. The causes are related not only to the high performance required in

models for disease diagnosis but also to the sensitivity of the data processed and the protection of patient

privacy. There are stringent policies on medical image dissemination to prevent image theft, image de-

anonymization, data poisoning attacks, and other security issues. The proposed system for AD diagnosis from

RGB MRI brain images implements the Federated Learning (FL) architecture and a strategy of medical image

authentication through biometric recognition to protect the privacy and confidentiality of the medical image

used for the training model and to mitigate the data poisoning attacks on the model. Experiments conducted

on two datasets of RGB MRI images (OASIS and ADNI) demonstrate that the proposed system achieved

performance comparable to a centralized ML system without a privacy-preserving strategy. The proposed

system represents a solution to solve security and privacy issues in a healthcare application for AD diagnosis.

1 INTRODUCTION

Alzheimer’s disease (AD) is a neurodegenerative

disease that affects the nerve cells within the brain,

progressively leading to irreversible cognitive

decline. Common symptoms of AD are to be found in

memory loss, cognitive decline, and changes in

personality and behavior. AD is one of the most

common neurodegenerative diseases, and it is

estimated that by 2050, one in 85 people will be

affected by AD (Wang et al., 2018). AD is incurable,

but an early diagnosis would allow a slower decline

of the disease, leading to significant benefits for

affected individuals (Angelillo, Balducci, et al., 2019;

Ghosh et al., 2023a). Several methodologies are used

to detect AD, such as Magnetic resonance imaging

(MRI), computer tomography (CT), positron

emission tomography (PET), and behavioral

biometrics through Human Activity Recognition

(HAR) (Angelillo, Impedovo, et al., 2019; Cicirelli et

al., 2022; Vincenzo Dentamaro et al., 2021; Gattulli

et al., 2022; Gattulli, Impedovo, Pirlo, & Sarcinella,

2023; Gattulli, Impedovo, Pirlo, & Semeraro, 2023;

https://orcid.org/0000-0002-8579-8941

https://orcid.org/0000-0002-9285-2555

https://orcid.org/0000-0002-7305-2210

Impedovo et al., 2012; Salehi et al., 2020). MRIs have

been increasingly used to diagnose AD through

machine learning (ML) methods. Among the ML

methods, Convolutional Neural Networks (CNN) are

the most promising and commonly used approaches

(Wen et al., 2020), (Cheriet et al., 2023). However,

using ML in healthcare involves several significant

risks to security and privacy (Sivan & Zukarnain,

2021). On the other hand, medical imbalanced data

affects the performance of ML diagnostic systems

(Vincenzo Dentamaro et al., 2018). For this reason,

these technologies have difficulty becoming a

standard in healthcare applications. Wrong or

improper use of these tools in healthcare would lead

to disastrous consequences that could even cost

people's lives. Some of the most critical issues related

to user privacy and security are constraining policies

on medical image dissemination, lack of data for

optimal model learning, interception attacks, de-

anonymization attacks, data poisoning attacks,

malware attacks, lack of transparency, and any type

of anomalies (Cannarile et al., 2022; Carrera et al.,

2022; V. Dentamaro et al., 2021).

Castro, F., Impedovo, D. and Pirlo, G.

A Federated Learning System with Biometric Medical Image Authentication for Alzheimer’s Diagnosis.

DOI: 10.5220/0012550200003654

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 13th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2024), pages 951-960

ISBN: 978-989-758-684-2; ISSN: 2184-4313

951

Federated learning (FL) represents a promising

technique to mitigate some of these issues (Li, Wen,

et al., 2021). FL enables decentralized learning by

preventing healthcare institutions from sharing their

data externally. The institutions participate in the

training of a global model and receive the knowledge

learned from the data of all institutions. The basic

idea is to train a local model with local data for each

healthcare institution and share only the local model

parameters. A global model receives and updates its

parameters with the local parameters of each

organization through different strategies. Finally, the

updated global model is shared with all the

organization's participants. Therefore, FL is a

privacy-preserving approach to compliance with

GDPR requirements and encourages healthcare

institutions to share knowledge acquired from their

data while maintaining confidentiality. However, the

FL approach is vulnerable to data poisoning attacks

and other types of attacks related to ML algorithms

(Mothukuri et al., 2021). Figure 1 shows some

possible attacks in the FL system for healthcare

applications. A data poisoning attack consists of

inserting images that cause the system to learn

incorrectly or, even worse, cause the system to give

the desired output. In (CHAN-HON-TONG, 2018) an

algorithm to perform an invisible data poisoning

attack that compromises the prediction result of a

deep learning model has been presented. Data

poisoning attacks could have disastrous

consequences in healthcare applications.

The proposed work presents an FL system to

diagnose AD from RGB MRI’s brain image,

implementing a method of authenticating MRI

images through physician biometric data to mitigate

data poisoning attacks. The proposed method

incorporates the fingerprint of the physician

authorized to give input images for local model

training within an RGB MRI image. The embedding

is done with a digital watermarking technique using

the Discrete Wavelet Transform (DWT) and the

Singular Value Decomposition (SVD) that ensures

the visual security image (Castro et al., 2023). The

physician’s fingerprint is extracted from the RGB

MRI image for authentication before training the

local model. In this way, only the authorized

physicians can train the local model. Subsequently,

the trained parameters of the local model are sent to

the global model for FL.

Few state-of-the-art approaches have explored the

use of FL in the context of AD diagnosis through

MRIs (Ghosh et al., 2023b; Khalil et al., 2023), and

none have addressed the problem of data poisoning

attacks. The main contributions of the proposed work

are:

• Develop an FL system to diagnose AD from

RGB MRI images, ensuring patient’s

privacy;

• Develop a methodology of RGB MRI image

authentication to mitigate data poisoning

attacks based on biometric recognition;

• Compare the proposed system performance

with no-FL system performance to evaluate

the trade-off between safety and

performance.

2 RELATED WORKS

Several studies have shown the effectiveness of ML

approaches for AD diagnosis from MRI images

(Mirzaei & Adeli, 2022). Many studies have

compared results obtained with deep learning

approaches with shallow learning approaches. Deep

learning techniques allow more accurate performance

in recognizing AD, especially in the case of different

stages of the disease (Suresha & Parthasarathy, 2020).

Among the most recent ML approaches include

enhanced probabilistic neural networks, neural

dynamic classification, dynamic ensemble learning,

and finite element machines. However, these

approaches do not protect the privacy of medical

images and patients. For this reason, FL systems have

been developed for healthcare applications. FL

approaches are used in contexts where data privacy is

an essential requirement. Kaissis et al. highlight the

great importance of using privacy protection

techniques in learning systems that process medical

images by presenting PriMIA (Kaissis et al., 2021). It

is an open-source framework that implements

learning neural networks on medical images by the

FL method and uses prevention techniques for a de-

anonymization and dataset reconstruction attack.

PriMIA's focus is to protect sensitive data at every

stage of learning while not sacrificing the

performance of shared learning. FL is used in the

healthcare domain for various tasks, such as brain

tumors, mammograms, issues due to COVID-19, and

more, as shown by Moon et al. (Narmadha &

Varalakshmi, 2022). Interest in FL systems has

involved institutions, government agencies, and

multinational companies specializing in tech and IT.

Nvidia has contributed to the release of Nvidia Flare,

an SDK focusing on artificial intelligence and FL. In

addition, Nvidia has released software specifically for

applications in the medical field with MONAI

NeroPRAI 2024 - Workshop on Medical Condition Assessment Using Pattern Recognition: Progress in Neurodegenerative Disease and

Beyond

952

Figure 1: Vulnerabilities in the FL system.

(Medical Open Network for AI), on which FL

systems can be implemented.

As well as a privacy-preserving approach, FL is

also used to solve the problem of fragmentation of

medical datasets. It results in low learning

performance and poor generalization of ML models.

The performance obtained with the FL approach

decentralized is similar to the performance of a

centralized model. It thus shows how FL is a possible

solution to the problem of dataset sparsity (Nguyen et

al., 2022).

FL is used in both supervised and unsupervised

learning. Bercea et al present FedDis (Bercea et al.,

2022), an FL framework that aims to greatly simplify

the work of clinicians by attempting to identify and

segment abnormalities in brain MRIs that are part of

the dataset. Labeling a dataset is arduous, and an

unsupervised Autoencoder neural network was used

to cluster the various MRIs produced by different

healthcare institutions through FL.

Only a few studies have focused on FL

approaches for AD recognition from MRI images

(Ghosh et al., 2023b; Zhao & Huang, 2023). The

accuracy and sensitivity of AD prediction with the FL

approach are comparable with traditional ML

approaches (Ghosh et al., 2023b).

Along with innumerable capabilities, FL has

some critical issues related to data heterogeneity and

ML attacks, such as reversal model and data

poisoning attacks. Homomorphic encryption (HE)

and secure multiparty computation (SMC) based on

secret sharing have been proposed against reversal

model attacks (Galantucci et al., 2021; Zhang et al.,

2023).

(Li, He, et al., 2021) focus their studies on the

issue of heterogeneous data used in FL, indicating

how the diversification of these can lead to poor

performance of the model. They propose a method for

locally correcting model updates by maximizing the

representation similarity between local and global

models. It has also been proposed as a valid method

for resisting data poisoning attacks. Data poisoning is

a relevant security issue in FL and ML algorithms.

Data poisoning attacks compromise the final result or

direct the model to a specific prediction, representing

a significant issue, especially in health diagnostic

systems. Han et al. propose a digital watermarking

technique to prevent manipulation of the test medical

images (Han et al., 2023). The technique is

appropriate for medical images because it does not

change important information in medical images and

ensures privacy data. Digital watermarking can

prevent image manipulation and ensure image

authentication for the model training phase.

The proposed work presents an FL system to

diagnose AD from RGB MRI images along with a

method of authenticating MRI images through

physician biometric data. Therefore, the proposed

system ensures the privacy of medical images and

their authentication against data poisoning attacks,

thus making the system robust for use in real-world

settings.

3 PROPOSED METHOD

The proposed system is based on the FL approach to

ensure the privacy of medical images used in the

training phase. The proposed system architecture is

shown in Figure 2. Each medical institution trains the

local model with local RGB MRI images. The local

RGB MRI images are authenticated with the

physician’s fingerprint before local training to

A Federated Learning System with Biometric Medical Image Authentication for Alzheimer’s Diagnosis

953

Figure 2: Proposed system architecture.

prevent data poisoning attacks. This stage is detailed

in Section 3.1. Subsequently, the authenticated MRI

images are used to train the local model, and the

training parameters are sent to the central server. The

central server aggregates the local parameters into a

global model through the FedAvg strategy and sends

the global model to each medical institution. The

communication between medical institutions and the

central server, the sending of parameters, the

aggregation strategy, and the updating of the global

model are detailed in Section 3.2. The architecture of

the local model used for AD diagnosis is described in

Section 3.3. The proposed system represents a

strategy to implement an ML system for AD

prediction with privacy protection and robustness to

data poisoning attacks.

3.1 Preprocessing and Image

Authentication Method

The preprocessing step includes embedding the

physician’s fingerprint to authenticate the MRI

through a digital watermarking technique. The

technique used is the Discrete Wavelet Transform

(DWT), which consists of dividing the RGB MRI

image into 4 sub-bands where each sub-band

represents a specific image detail: horizontal (LH),

vertical (HL), diagonal (HH), and approximation

(LL). The LL sub-band is the coefficient matrix that

contains the main image data. Figure 3 shows an RGB

MRI image divided into the 4 sub-band through the

DWT.

The LL sub-band cannot be changed, while the

remaining sub-bands can be changed and replaced

with other values to provide a visual secure

embedding. The embedding of the physician’s

fingerprint is performed by replacing the values of

LH, HL, and HH with the fingerprint image

factorization values. For the image factorization is

used the Singular Value Decomposition (SVD)

divides the image into 2 orthogonal matrixes U and

VT) and 1 vector of singular values (S). Both DWT

and SVD are common techniques of digital

watermarking because they consist of invertible

transformations. Therefore, it is possible to

reconstruct the original image from the values

obtained by transformations. The sub-band LH and

HL are replaced with U and VT matrixes respectively.

The values of vector S are inserted into the HH sub-

band following through substitutions and

permutations to introduce entropy into embedding

and make it complex for an attacker to reconstruct the

fingerprint image. Finally, the RGB MRI image with

the embedding of the fingerprint image (watermarked

image) is generated through the Inverse Discrete

Wavelet Transform (IDWT) given as input LL and

the replaced sub-bands. Figure 3 shows the results of

the IDWT.

Figure 4 shows the visual security of the proposed

method. The proposed method enables biometric

authentication to be applied to the medical image

used for model training. Therefore, the physician’s

fingerprint image is extracted from a watermarked

image to authenticate the image before starting local

model training, avoiding training with poisoned

images. The extraction process involves applying

DTW on the watermarked image to re-obtain the 4

sub-bands (LL’, LH’, HL’, and HH’). Consequently,

LH’, HL’, and HH’ contain the values of U, VT, and

NeroPRAI 2024 - Workshop on Medical Condition Assessment Using Pattern Recognition: Progress in Neurodegenerative Disease and

Beyond

954

Figure 3: a) Original MRI image. b) LL of MRI image. c) LH of MRI image. d) HL of MRI image. e) HH of MRI image.

S respectively, used to reconstruct the fingerprint

image through the SVD reconstruction. It consists of

converting S into a diagonal matrix (D), and then the

dot product between U, D, and VT is performed to

obtain the image.

Figure 4: (a) original MRI image (b) fingerprint image (c)

MRI image with embedding fingerprint (watermarked

image).

3.2 Federated Learning

The proposed system is based on an FL architecture

in which each medical institution participates in

training the global model without sharing its patients'

sensitive data. The entire process consists of 5 steps,

as described below.

• Step 1: actors in the entire process are described,

namely, the server and the various clients or

nodes. The nodes participating in the federation

are the hospitals or medical institutions that

possess a database of RGB MRI images needed

for AD prediction. The server owns the model to

be trained that will be shared with the various

participating nodes.

• Step 2: the server synchronizes with all nodes,

sending the model base with null or random

weights.

• Step 3: each node performs training of the newly

received model with its local dataset of RGB

MRI images.

• Step 4: the nodes only send the weights obtained

during the learning phase to the server. The

server performs the FedAvg aggregation

approach of all the weights received to create a

new model representing the learning performed

by all the various nodes. The FedAvg combines

the local gradient computation assigned to each

node with the gradient averaging computation

performed by the server. On the other hand, the

received weights are weighted according to the

number of training examples of each node, and

then the weights averaging is performed. FedAvg

is robust to unbalanced and non-IDD

distributions.

• Step 5: the server synchronizes with each node

by sending the new version of the model and

iteratively restarts the entire FL process.

The entire FL process is iterated several times by

updating the global model at each iteration to allow

the global model to learn the knowledge obtained

from each node.

3.3 Convolutional Neural Network

(CNN)

The model used for the training is based on a CNN,

which is able to process RGB MRI images to detect

significant patterns of the presence of AD. CNN has

special convolutional layers, which can apply filters

to the original image to recognize and extract patterns

characteristic of the image. At a low level, thus

considering the first few layers, the CNN applies

convolutional filters to be able to recognize essential

elements such as lines, angles, figures, etc., and then

goes on to more and more complex elements that, at

a high level, might be characteristic of those suffering

from cognitive impairment due to AD such as more

irregular and less defined grooves. To reduce the

computational cost and choose the features with

higher magnitude, CNN has layers for the Max

Pooling technique, which consists of choosing the

coefficient with a higher value within a submatrix of

the input. This reduces the dimensionality of the

image and favors the most significant features. The

extracted features are given as input to a Flatten layer

to reduce the dimensionality of the images. The last

output layer is represented by a Dense layer of binary

classification (disease or no disease), consisting of a

single neuron with a sigmoidal activation function.

A Federated Learning System with Biometric Medical Image Authentication for Alzheimer’s Diagnosis

955

Summarizing the implemented CNN architecture is

constituted by:

• two-dimensional convolutional input layer

with Relu activation function;

• two Max Pooling layers;

• flatten layer;

• dense layer with a sigmoid activation

function.

4 EXPERIMENTS

The proposed system is tested using two datasets of

RGB MRI images to predict AD, as described in

Section 4.1. Two scenarios have been simulated to

evaluate the effectiveness of the proposed system

compared to a traditional ML system. The first

scenario implements a traditional ML system based

on the CNN described in Section 3.3 where the

medical institutions share the local RGB MRI images

into a central database to train the model for AD

prediction. This scenario presents vulnerabilities

related to data privacy and data poisoning attacks.

The second scenario implements the proposed system

with FL architecture and RGB MRI image

authentication. This scenario is privacy-preserving

and robust to data poisoning attacks. The

performances obtained in both scenarios are

evaluated in terms of accuracy, precision, recall, and

F1-scores for healthy and sick classes. The results are

compared to evaluate the cost of implementing the

proposed system. The CNN model in both scenarios

is trained with the same parameters for a correct

comparison. Specifically, the model has been trained

with 50 epochs and batch size 32 using the ‘early

stopping’ approach with the patience of 5 epochs to

prevent overfitting. The results obtained in the first

and second scenarios and their comparison have been

detailed in Sections 4.2, 4.3, and 4.4, respectively.

4.1 Datasets

For the experiments, OASIS and ADNI datasets were

used. Open Access Series of Imaging Studies

(OASIS) (Marcus et al., 2007) dataset is a cross-

sectional RGB MRI image collection of 416 subjects

aged between 18 and 96 years. One hundred of these

subjects older than 60 years have been clinically

diagnosed with very mild to moderate AD. The

subjects include both men and women. For each

subject, 3 or 4 individual T1-weighted MRI scans

obtained in single scan sessions are included. Data-

augmentation techniques based on image

manipulation have been used to solve the dataset

imbalance. Techniques perform slight editing

operations on an image, including rotations,

extensions, and compressions. At the end of the data-

augmentation technique, a balanced dataset of 526

RGB MRI images of healthy subjects and 526 RGB

MRI images of subjects with AD has been obtained.

The Alzheimer’s Disease Neuroimaging Initiative

(ADNI) dataset (Jack et al., 2008) includes 6400

cross-sectional RGB MRI images of 3200 healthy and

3200 subjects with AD. The images represent. The

dataset is balanced, and then no data-augmentation

technique is applied.

4.2 Results in Traditional ML Scenario

The first experiment consists of testing the CNN

performance in the no-FL scenario. The datasets are

divided into train and test sets using an 80 - 20 ratio.

The results are reported in Table 1.

Figure 5: a) Confusion matrix of OASIS dataset. b)

confusion matrix of the ADNI dataset.

NeroPRAI 2024 - Workshop on Medical Condition Assessment Using Pattern Recognition: Progress in Neurodegenerative Disease and

Beyond

956

Table 1: Results of CNN model in no-FL scenario.

Metrics OASIS ADNI

Accurac

91.94% 90.70%

Precision 98.06% 93.47%

Recall 86.32% 87.88%

F1-score 92.94% 90.59%

The results for each healthy and AD class are reported

in the confusion matrix in Figure 5. The CNN model

shows the best performance in healthy subject

recognition with an accuracy of healthy subject

recognition of 99.02% and 93.47% for the OASIS and

ADNI datasets, respectively. The accuracy in AD

recognition is 85.32% and 88.16% for OASIS and

ADNI datasets, respectively.

4.3 Results of the Proposed System

The second experiment tests the proposed FL with an

image authentication system. Two clients that represent

two medical institutions and 5 rounds of learning are

used for the FL simulation. Each round represents a

single iteraction of global model update as discussed in

Section 3.2. Each client implements the same CNN

model, and the datasets are divided equally among

clients. Specifically, the training set for each dataset is

split 50-50 between the two clients, preserving the

balance of the classes. The proposed image

authentication approach is applied to each training

image, and then the fingerprint image is embedded into

each RGB MRI image before model training.

Subsequently, the fingerprint image has been extracted

and the RGB MRI image without the fingerprint is

given as input to CNN. The proposed simulation is

necessary to thoroughly validate the proposed system

because the proposed image authentication technique

inevitably reduces RGB MRI images quality.

Therefore, the experiments have a dual purpose of

evaluating how the image quality reduction impacts the

model performance and evaluating the impact of the FL

strategy. The results in terms of accuracy for each

learning round are reported in Table 2.

Table 2: Accuracy results of the proposed system.

OASIS ADNI

Round 1 89.20% 82.06%

Round 2 89.61% 84.64%

Round 3 89.61% 85.34%

Round 4 92.00% 86.98%

Round 5 92.00% 88.50%

Table 2 shows that the proposed system improves

the performance at each iteration until the optimal

performance of the global model is obtained at the

fifth iteration with 92% and 88.50% accuracy for the

OASIS and ADNI datasets, respectively.

4.4 Comparison of Results

The results obtained with the traditional and the

proposed system in terms of accuracy are compared

in Figure 6 and Figure 7 for the OASIS and ADNI

datasets, respectively. Figures 6 and 7 show how the

proposed system's accuracy matches the traditional

system's accuracy during the 5 rounds of FL. The

results of the final FL models are comparable with the

results of the traditional system, with an accuracy

degradation of 2.20% in the ADNI dataset and an

accuracy improvement of 0.06% in the case of the

OASIS dataset. Therefore, the proposed system

ensures performance comparable with a traditional

system, ensuring privacy and mitigating data

poisoning attacks. The system is efficient on both

tested datasets that are heterogeneous in data

quantity, image resolution, and scanning perspective.

Therefore, the system is flexible in processing and

learning multiple types of information of different

quantities and qualities.

Figure 6: Comparison of results between FL and no FL

approach for the OASIS dataset.

Figure 7: Comparison of results between FL and no FL

approach for the ADNI dataset.

100

ROUND

FL Avg Accuracy No FL Avg Accuracy

100

ROUND

FL Avg Accuracy No FL Avg Accuracy

A Federated Learning System with Biometric Medical Image Authentication for Alzheimer’s Diagnosis

957

5 CONCLUSIONS

A system based on FL environment and MRI images

authentication to predict AD from RGB MRI brain

images has been proposed. The proposed system

combines the potential of machine learning systems

with protecting the privacy of sensitive data and

patient information and is robust to data poisoning

attacks that could cause disastrous consequences in a

healthcare application. The proposed system is a

solution to introduce AD recognition in a healthcare

context to support physicians. The proposed system

is compliance with stringent requirements of privacy

and security in GDPR and other international

regulations. Experiments have shown that the

proposed system has a good trade-off between safety

and performance. The accuracy in AD recognition is

degraded by 2.20% with the ADNI dataset and it is

improved by 0.06% with the OASIS dataset using the

proposed system. The proposed system is tested,

simulating an FL scenario with two medical

institutions. In the future, the proposed system could

be tested with a large number of medical institutions

using a more significant number of MRI images.

Moreover, multimodal AD recognition could be

implemented using MRI images with different

scanning perspectives and other medical information.

The proposed system is privacy-preserving and

prevents data poisoning attacks but there are other

potential security issues in FL, such as model reversal

attacks and compromise attacks on the global model

that have not been considered. In the future, the

security of the proposed system could be improved by

implementing defense strategies against these attacks.

ACKNOWLEDGEMENTS

Francesco Castro is a PhD student enrolled in the

National PhD in Artificial Intelligence, XXXVIII

cycle, course on Health and life sciences, organized

by Università Campus Bio-Medico di Roma.

REFERENCES

Angelillo, M. T., Balducci, F., Impedovo, D., Pirlo, G., &

Vessio, G. (2019). Attentional Pattern Classification for

Automatic Dementia Detection. IEEE Access, 7,

57706–57716. https://doi.org/10.1109/ACCESS.2019.

2913685

Angelillo, M. T., Impedovo, D., Pirlo, G., & Vessio, G.

(2019). Performance-Driven Handwriting Task

Selection for Parkinson’s Disease Classification.

Lecture Notes in Computer Science (Including

Subseries Lecture Notes in Artificial Intelligence and

Lecture Notes in Bioinformatics), 11946 LNAI, 281–

293. https://doi.org/10.1007/978-3-030-35166-3_20/

COVER

Bercea, C. I., Wiestler, B., Rueckert, D., & Albarqouni, S.

(2022). Federated disentangled representation learning

for unsupervised brain anomaly detection. Nature

Machine Intelligence 2022 4:8, 4(8), 685–695.

https://doi.org/10.1038/s42256-022-00515-2

Cannarile, A., Dentamaro, V., Galantucci, S., Iannacone,

A., Impedovo, D., & Pirlo, G. (2022). Comparing Deep

Learning and Shallow Learning Techniques for API

Calls Malware Prediction: A Study. Applied Sciences

2022, Vol. 12, Page 1645, 12(3), 1645.

https://doi.org/10.3390/APP12031645

Carrera, F., Dentamaro, V., Galantucci, S., Iannacone, A.,

Impedovo, D., & Pirlo, G. (2022). Combining

Unsupervised Approaches for Near Real-Time

Network Traffic Anomaly Detection. Applied Sciences

2022, Vol. 12, Page 1759, 12(3), 1759.

https://doi.org/10.3390/APP12031759

Castro, F., Impedovo, D., & Pirlo, G. (2023). A Medical

Image Encryption Scheme for Secure Fingerprint-

Based Authenticated Transmission. Applied Sciences

2023, Vol. 13, Page 6099, 13(10), 6099.

https://doi.org/10.3390/APP13106099

CHAN-HON-TONG, A. (2018). An Algorithm for

Generating Invisible Data Poisoning Using Adversarial

Noise That Breaks Image Classification Deep Learning.

Machine Learning and Knowledge Extraction 2019,

Vol. 1, Pages 192-204, 1(1), 192–204.

https://doi.org/10.3390/MAKE1010011

Cheriet, M., Dentamaro, V., Hamdan, M., Impedovo, D., &

Pirlo, G. (2023). Multi-speed transformer network for

neurodegenerative disease assessment and activity

recognition. Computer Methods and Programs in

Biomedicine, 230. https://doi.org/10.1016/J.CMPB.

2023.107344

Cicirelli, G., Impedovo, D., Dentamaro, V., Marani, R.,

Pirlo, G., & D’Orazio, T. R. (2022). Human

Gait Analysis in Neurodegenerative Diseases: A

Review. IEEE Journal of Biomedical and Health

Informatics, 26(1), 229–242. https://doi.org/10.1109/

JBHI.2021.3092875

Dentamaro, V., Convertini, V. N., Galantucci, S., Giglio,

P., Palmisano, T., & Pirlo, G. (2021). Ensemble

consensus: An unsupervised algorithm for anomaly

detection in network security data. CEUR WORKSHOP

PROCEEDINGS, 2940, 309–318. https://ricerca.

uniba.it/handle/11586/377597

Dentamaro, Vincenzo, Impedovo, D., & Pirlo, G. (2018).

LICIC: Less Important Components for Imbalanced

Multiclass Classification. Information 2018, Vol. 9,

Page 317, 9(12), 317. https://doi.org/10.3390/

INFO9120317

Dentamaro, Vincenzo, Impedovo, D., & Pirlo, G. (2021).

An Analysis of Tasks and Features for Neuro-

Degenerative Disease Assessment by Handwriting.

Lecture Notes in Computer Science (Including

NeroPRAI 2024 - Workshop on Medical Condition Assessment Using Pattern Recognition: Progress in Neurodegenerative Disease and

Beyond

958

Subseries Lecture Notes in Artificial Intelligence and

Lecture Notes in Bioinformatics), 12661 LNCS, 536–

545. https://doi.org/10.1007/978-3-030-68763-2_41/

COVER

Galantucci, S., Impedovo, D., & Pirlo, G. (2021). One time

user key: A user-based secret sharing XOR-ed model

for multiple user cryptography in distributed systems.

IEEE Access, 9, 148521–148534. https://doi.

org/10.1109/ACCESS.2021.3124637

Gattulli, V., Impedovo, D., Pirlo, G., & Sarcinella, L.

(2023). Human Activity Recognition for the

Identification of Bullying and Cyberbullying Using

Smartphone Sensors. Electronics 2023, Vol. 12, Page

261, 12(2), 261. https://doi.org/10.3390/

ELECTRONICS12020261

Gattulli, V., Impedovo, D., Pirlo, G., & Semeraro, G.

(2022). Early Dementia Identification: On the Use of

Random Handwriting Strokes. Lecture Notes in

Computer Science (Including Subseries Lecture Notes

in Artificial Intelligence and Lecture Notes in

Bioinformatics), 13424 LNCS, 285–300.

https://doi.org/10.1007/978-3-031-19745-1_21

Gattulli, V., Impedovo, D., Pirlo, G., & Semeraro, G.

(2023). Handwriting Task-Selection based on the

Analysis of Patterns in Classification Results on

Alzheimer Dataset. CEUR Workshop Proceedings,

3521, 18–29.

Ghosh, T., Palash, M. I. A., Yousuf, M. A., Hamid, M. A.,

Monowar, M. M., & Alassafi, M. O. (2023a). A Robust

Distributed Deep Learning Approach to Detect

Alzheimer’s Disease from MRI Images. Mathematics

2023, Vol. 11, Page 2633, 11(12), 2633.

https://doi.org/10.3390/MATH11122633

Ghosh, T., Palash, M. I. A., Yousuf, M. A., Hamid, M. A.,

Monowar, M. M., & Alassafi, M. O. (2023b). A Robust

Distributed Deep Learning Approach to Detect

Alzheimer’s Disease from MRI Images. Mathematics

2023, Vol. 11, Page 2633, 11(12), 2633.

https://doi.org/10.3390/MATH11122633

Han, B., Jhaveri, R. H., Wang, H., Qiao, D., & Du, J.

(2023). Application of Robust Zero-Watermarking

Scheme Based on Federated Learning for Securing the

Healthcare Data. IEEE Journal of Biomedical and

Health Informatics, 27(2), 804–813. https://doi.

org/10.1109/JBHI.2021.3123936

Impedovo, D., Pirlo, G., Sarcinella, L., Stasolla, E., &

Trullo, C. A. (2012). Analysis of stability in static

signatures using cosine similarity. Proceedings -

International Workshop on Frontiers in Handwriting

Recognition, IWFHR, 231–235. https://doi.org/10.

1109/ICFHR.2012.180

Jack, C. R., Bernstein, M. A., Fox, N. C., Thompson, P.,

Alexander, G., Harvey, D., Borowski, B., Britson, P. J.,

Whitwell, J. L., Ward, C., Dale, A. M., Felmlee, J. P.,

Gunter, J. L., Hill, D. L. G., Killiany, R., Schuff, N.,

Fox-Bosetti, S., Lin, C., Studholme, C., … Weiner, M.

W. (2008). The Alzheimer’s Disease Neuroimaging

Initiative (ADNI): MRI methods. Journal of Magnetic

Resonance Imaging : JMRI

, 27(4), 685–691.

https://doi.org/10.1002/JMRI.21049

Kaissis, G., Ziller, A., Passerat-Palmbach, J., Ryffel, T.,

Usynin, D., Trask, A., Lima, I., Mancuso, J., Jungmann,

F., Steinborn, M. M., Saleh, A., Makowski, M.,

Rueckert, D., & Braren, R. (2021). End-to-end privacy

preserving deep learning on multi-institutional medical

imaging. Nature Machine Intelligence 2021 3:6, 3(6),

473–484. https://doi.org/10.1038/s42256-021-00337-8

Khalil, K., Khan Mamun, M. M. R., Sherif, A., Elsersy, M.

S., Imam, A. A. A., Mahmoud, M., & Alsabaan, M.

(2023). A Federated Learning Model Based on

Hardware Acceleration for the Early Detection of

Alzheimer’s Disease. Sensors 2023, Vol. 23, Page

8272, 23(19), 8272. https://doi.org/10.3390/S23198272

Li, Q., He, B., & Song, D. (2021). Model-Contrastive

Federated Learning. 2021 IEEE/CVF Conference on

Computer Vision and Pattern Recognition (CVPR),

10708–10717. https://doi.org/10.1109/CVPR46437.

2021.01057

Li, Q., Wen, Z., Wu, Z., Hu, S., Wang, N., Li, Y., Liu, X.,

& He, B. (2021). A Survey on Federated Learning

Systems: Vision, Hype and Reality for Data Privacy

and Protection. IEEE Transactions on Knowledge and

Data Engineering. https://doi.org/10.1109/TKDE.

2021.3124599

Marcus, D. S., Wang, T. H., Parker, J., Csernansky, J. G.,

Morris, J. C., & Buckner, R. L. (2007). Open Access

Series of Imaging Studies (OASIS): Cross-sectional

MRI Data in Young, Middle Aged, Nondemented, and

Demented Older Adults. Journal of Cognitive

Neuroscience, 19(9), 1498–1507. https://doi.org/10.

1162/JOCN.2007.19.9.1498

Mirzaei, G., & Adeli, H. (2022). Machine learning

techniques for diagnosis of alzheimer disease, mild

cognitive disorder, and other types of dementia.

Biomedical Signal Processing and Control, 72, 103293.

https://doi.org/10.1016/J.BSPC.2021.103293

Mothukuri, V., Parizi, R. M., Pouriyeh, S., Huang, Y.,

Dehghantanha, A., & Srivastava, G. (2021). A survey

on security and privacy of federated learning. Future

Generation Computer Systems, 115, 619–640.

https://doi.org/10.1016/J.FUTURE.2020.10.007

Narmadha, K., & Varalakshmi, P. (2022). Federated

Learning in Healthcare: A Privacy Preserving

Approach. Studies in Health Technology and

Informatics, 294, 194–198. https://doi.org/10.3233/

SHTI220436

Nguyen, T. V., Dakka, M. A., Diakiw, S. M., VerMilyea,

M. D., Perugini, M., Hall, J. M. M., & Perugini, D.

(2022). A novel decentralized federated learning

approach to train on globally distributed, poor quality,

and protected private medical data. Scientific Reports

2022 12:1, 12(1), 1–12. https://doi.org/10.1038/

s41598-022-12833-x

Salehi, A. W., Baglat, P., Sharma, B. B., Gupta, G., &

Upadhya, A. (2020). A CNN Model: Earlier Diagnosis

and Classification of Alzheimer Disease using MRI.

Proceedings - International Conference on Smart

Electronics and Communication, ICOSEC 2020, 156–

161. https://doi.org/10.1109/ICOSEC49089.2020.9215

402

A Federated Learning System with Biometric Medical Image Authentication for Alzheimer’s Diagnosis

959

Sivan, R., & Zukarnain, Z. A. (2021). Security and Privacy

in Cloud-Based E-Health System. Symmetry 2021, Vol.

13, Page 742, 13(5), 742. https://doi.org/

10.3390/SYM13050742

Suresha, H. S., & Parthasarathy, S. S. (2020). Alzheimer

Disease Detection Based on Deep Neural Network with

Rectified Adam Optimization Technique using MRI

Analysis. Proceedings of 2020 3rd International

Conference on Advances in Electronics, Computers and

Communications, ICAECC 2020. https://doi.org/10.

1109/ICAECC50550.2020.9339504

Wang, S. H., Phillips, P., Sui, Y., Liu, B., Yang, M., &

Cheng, H. (2018). Classification of Alzheimer’s

Disease Based on Eight-Layer Convolutional Neural

Network with Leaky Rectified Linear Unit and Max

Pooling. Journal of Medical Systems, 42(5), 1–11.

https://doi.org/10.1007/S10916-018-0932-

7/FIGURES/11

Wen, J., Thibeau-Sutre, E., Diaz-Melo, M., Samper-

González, J., Routier, A., Bottani, S., Dormont, D.,

Durrleman, S., Burgos, N., & Colliot, O. (2020).

Convolutional neural networks for classification of

Alzheimer’s disease: Overview and reproducible

evaluation. Medical Image Analysis, 63, 101694.

https://doi.org/10.1016/J.MEDIA.2020.101694

Zhang, L., Xu, J., Vijayakumar, P., Sharma, P. K., &

Ghosh, U. (2023). Homomorphic Encryption-Based

Privacy-Preserving Federated Learning in IoT-Enabled

Healthcare System. IEEE Transactions on Network

Science and Engineering, 10(5), 2864–2880.

https://doi.org/10.1109/TNSE.2022.3185327

Zhao, L., & Huang, J. (2023). A distribution information

sharing federated learning approach for medical image

data. Complex and Intelligent Systems, 9(5), 5625–

5636. https://doi.org/10.1007/S40747-023-01035-

1/FIGURES/11.

NeroPRAI 2024 - Workshop on Medical Condition Assessment Using Pattern Recognition: Progress in Neurodegenerative Disease and

Beyond

960