Advancements and Applications of Federated Learning in Biometric

Recognition

Zhengliang Lyu

Computer Science and Technology, China University of Mining and Technology-Beijing, Beijing, China

Keywords: Federated Learning, Biometrics, Data Privacy Protection.

Abstract: The extensive use of biometric technologies has rendered the enhancement of model performance, while

safeguarding user privacy, a significant concern. Federated learning, an emerging distributed machine

learning technique, enhances model generalization and accuracy while safeguarding user privacy through

collaborative training across several devices. This paper reviews the application progress of federated learning

in the field of biometrics, and discusses its advantages in improving model performance and protecting user

privacy, as well as the challenges and future development directions. Specifically, this paper first introduces

different types of biometrics in the past and their disadvantages, the basic concepts and advantages of

federated learning. It subsequently conducts a detailed analysis of the application of federated learning in

biometrics, encompassing data diversity, real-time processing, and privacy protection. Then, this paper

discusses the challenges faced by federated learning in biometrics, such as data leakage risk, high computing

resource demand and uneven data distribution, and proposes optimization strategies such as edge cloud

collaborative computing and distributed computing optimization. Finally, this article anticipates future

advancements in federated learning within biometrics. The significance of this paper is to introduce the broad

prospects of federated learning in the field of biometrics, and provide an effective method to protect user

privacy for future research in biometrics

1 INTRODUCTION

Internet technologies and artificial intelligence have

pushed society toward comprehensive

Informatization and intelligence. This tendency has

made it simpler to transfer personal information

quickly, making people more aware of the need to

protect their privacy. Recent advances in artificial

intelligence have led to the widespread use of user

privacy data for deep learning model analysis in

recommendation, risk assessment, and identity

authentication systems. These technologies collect a

lot of user data, which researchers mine, analyze, and

choose from. Despite its benefits, technology is a

"double-edged sword," with large data analysis risks

privacy data leaks. When assessing customer

purchasing preferences on e-commerce platforms,

insufficient data protection may expose purchase

records and browsing histories. This may make

targeted fraud easier for crooks. Corporations must

also manage huge volumes of user identifying data.

https://orcid.org/0009-0003-3569-8349

Any data breach in identity authentication systems

could have dire consequences for users. As a result,

disclosing information violates user privacy and

causes further problems, which increases demand for

better privacy protection.

The fingerprint recognition technique is the best-

known biometric recognition system. Early on in the

1990s of the 20th century, various businesses sought

to fit fingerprint identification on mobile phones,

Siemens exploited Bromba's technology, and created

the first prototype mobile phone with fingerprint

recognition in 1998. But using this technology calls

for particular motions, which is a little uncomfortable.

It is only in recent years that capacitive, optical and

ultrasonic methods have begun to be widely used in

fingerprint recognition. Capacitive fingerprint

recognition is the most widely used capacitive

fingerprint recognition solution. The fingerprint

recognition sensor is combined with the subcutaneous

electrolyte of the human body to form an electric

field, and the unevenness of the fingerprint surface

526

Lyu and Z.

Advancements and Applications of Federated Learning in Biometric Recognition.

DOI: 10.5220/0013527700004619

In Proceedings of the 2nd International Conference on Data Analysis and Machine Learning (DAML 2024), pages 526-531

ISBN: 978-989-758-754-2

will cause the voltage difference between the two to

fluctuate in a certain amplitude. With this change, the

accuracy of fingerprint recognition will be improved.

Optical fingerprint recognition primarily illuminates

fingerprints using an integrated light source, gathers

reflected and refracted light, and subsequently directs

this light through a prism to illuminate the target

object. It analyzes the intensity and angle of the

reflected light, generating a graph with multiple gray

values, which is then compared against a pre-existing

database for matching (Li, 2020).

Palmprint recognition is a biometric identification

technology that offers better privacy protection

features. The most important characteristic of palm

prints is the features of the palm lines (Chen et al.,

2019). The main lines are the most stable and clearest

features among the palm prints, which can still be

distinctly identified even in low-resolution and low-

quality images. Huang et al. (Huang et al., 2008)

proposed a method for palm print verification based

on principal lines, using the Modified Finite Radon

Transform (MFRAT) to extract the principal lines of

the palm print. Experimental results demonstrate that

the principal lines have strong distinguishability. Jia

et al. (Jia et al., 2009) proposed a method for palm

print retrieval based on principal lines, detecting a

large number of key points on the three principal lines

and using these key points for retrieval. Li et al. (Li et

al., 2010) enhanced the quality of palm print images

using grayscale adjustment and median filtering

based on the characteristics of the main lines in the

palm print. They then detected the palm lines based

on diversity and contrast, refined the results using an

improved Hilditch algorithm, and finally obtained a

single-pixel palm print main line image using an edge

tracing method. The iris, as a biometric feature of the

human body, is difficult to replicate and has natural

anti-counterfeiting properties; it remains stable

throughout a person's life. In the early 1990s,

Daugman first proposed algorithmic architecture for

iris recognition systems (Daugman, 1993), and

researcher John Daugman from Cambridge

University pioneered the iris biometric identification

system for human recognition (Daugman, 1998).

Dougman designed the famous Integro-Differential

Operator (IDO) to mark the contours of the iris.

Firstly, preprocess the iris image to enhance contrast;

then, use differential operators (such as the Sobel

operator) to extract edge information and identify the

contours of the iris. Next, apply integral operators to

smooth the image and reduce noise impact; finally,

extract feature points and label them. This method can

effectively improve the accuracy and stability of iris

recognition.

Although these biometric technologies are already

in use, they have certain drawbacks: iris recognition

is very accurate, but its widespread application is

limited due to the high cost of establishing iris

recognition systems. The process of fingerprint

recognition often involves physical contact, making it

easy to be stolen in public places. These technologies

do not provide any form of comprehensive security

protection for personal privacy.

The remainder of the chapter is organized as

follows. Firstly, this study will introduce the method

of federated learning-based biometrics recognition in

Section 2. Then, in Section 3, the future prospects and

future challenges in the field of federated learning-

based biometrics recognition in general will be

discussed. Finally, Section 4 summarizes the paper

and gives the conclusions drawn from the discussion.

2 METHOD

2.1 Concept of Federated Learning

Problems with data privacy, user rights to use their

data, and the presence of data silos led to the

development of the federated learning idea. Federated

learning is a subset of distributed training that ensures

data is both invisible and accessible when training

deep learning models. It does not need collecting all

data on a single server. This method successfully

protects the private information of users while

simultaneously monitoring the model's progress in

training. By utilizing federated learning, any person

may process data locally and safely share model

updates with other participants at the same time. This

method protects data privacy while also doing away

with the hazards connected to centralized data storage.

Federated learning is a promising solution for striking

a balance between data utilization and privacy

protection, which will eventually benefit individuals

and society.

2.2 Federated Learning-based Face

Recognition

Federated learning, which keeps users' facial

information local to improve model generalization

while protecting privacy, has been introduced as a

means of privacy protection in recent years due to the

rapid development of facial recognition technology.

Due to variables like device capabilities and data

volume, the global optimized model that the federated

learning central server obtains may not always be the

local optimum for the participants. As such, getting a

Advancements and Applications of Federated Learning in Biometric Recognition

527

local model that works for each user is just as

important as taking into account the best global

model. Federated Face Recognition (FedFace)

(Aggarwal et al., 2021) was proposed by Aggarwal et

al. as a framework for group face recognition model

learning. Using the local data set, the client trains the

model updates before uploading them to the server.

FedFace uses federated learning to expand the data

set size and improve the performance of the facial

recognition model by leveraging the models that have

been trained on facial images from various clients.

However, this approach may also result in user

privacy violations because it requires uploading each

local model's class embedding matrix.

It was suggested by Meng et al. (Meng et al., 2022)

to include differential privacy technology in the

model update procedure. When mapping user facial

features onto a hypersphere during the model

aggregation process in traditional federated learning

frameworks, feature overlap can occur. This reduces

the recognition accuracy of the model. As a result,

Meng et al. (Meng et al., 2022) shares an embedding

matrix with other users using a de-identified form.

The server sends users a class center matrix with

noise perturbations from other users based on the

assumption that the server is trustworthy. Both

feature privacy and feature overlap are addressed by

the differential privacy local clustering algorithm

(DPLC).

Another problem with traditional federated

learning methods is that it becomes hard to compute

the negative loss that is used to distinguish between

different categories because a single client only has

access to data from one category and lacks category

information from other clients (Zhang et al., 2022).

This reduces the ability to distinguish between

category vectors, prevents inter-class distances from

growing, and consequently has an impact on the

model's overall performance. This paper offers a

novel solution to the previously mentioned problems,

which enables each client to calculate the negative

item loss without directly disclosing the category

vector information to other clients. This is

accomplished by creating a new, comprehensive

category vector on the server side by combining all of

the client-provided category vector data. The client

can compute the negative loss based on this newly

generated vector once it has been received. The

privacy and security of the data are guaranteed by the

way this new vector is made, which prevents it from

being used to retrieve any original category of vector

information from the client. This new fusion vector is

referred to as the aggregated category vector in this

article. This paper uses cosine similarity to calculate

the similarity of the aggregated category vectors after

they are generated, so that the aggregated category

vectors sent to the client cover a wider feature space.

The degree of direction similarity between two

vectors is measured using cosine similarity. The two

vectors are more similar when the cosine value is

closer to 1, which also indicates that the angle

between them is closer to 0 degrees. Thus, the

aggregated category vectors are sorted and the types

of aggregated category vectors are chosen based on

the computed cosine similarity values (Gao, 2024).

2.3 Fingerprint and Finger Vein

Recognition

2.3.1 Fingerprint Recognition Algorithm

A novel fingerprint recognition algorithm, termed

FedGFR, has been proposed to address the challenges

of subpar raw data quality and equitable client

selection in federated learning for fingerprint

recognition. The comprehensive system architecture

of FedGFR integrates numerous local models via

federated learning to develop a universal recognizer

while safeguarding user privacy on the client side. A

client fair scheduling strategy has been proposed to

achieve a more equitable distribution of accuracy

among clients in a federated learning framework.

Upon the arrival of new clients in the training set, it

is essential to determine whether to incorporate them

into the existing sample array, followed by the

random overwriting of sample data at a designated

position within the set. Clients develop their

recognition models through a federated learning

approach utilizing their local data. The server

acquires model parameters from data proprietors and

consolidates them, revising the global model

parameters via the FedAvg algorithm within the

federated server. The consolidated global model is

subsequently prior to the commencement of the new

iteration, equitably selecting the clients participating

in each round until the model reaches convergence.

Thus, privacy concerns associated with centralized

learning in conventional machine learning techniques

can be avoided (Wang, 2024).

2.3.2 Development of a Finger Vein

Recognition Algorithm N-Model

Personalized Federated Learning

In N-model-based personalized federated learning, a

series of aggregation algorithms is performed for N

clients at the central server, producing a distinct

aggregation model for each client on the server. Upon

DAML 2024 - International Conference on Data Analysis and Machine Learning

528

gathering the local models from all clients, the central

server will execute various model aggregation

algorithms for each client. Consequently, when the

finger vein datasets exhibit significant heterogeneity,

the algorithm proposed in this paper can produce

tailored aggregation models for each client on the

central server, thereby personalizing the local models

and effectively accommodating the distribution of

their local datasets (Lian, 2023).

3 DISCUSSIONS

3.1 Issues in Federated Learning-based

Biometrics Recognition

Federated learning has great potential in development

and application, but it also confronts tremendous

obstacles as shown below.

3.1.1 Computing Power Issues

Although federated learning can effectively improve

model performance by training multiple devices

together while protecting the privacy of all parties, the

computational power of today's mobile devices only

allows for the implementation of algorithms with

small computational loads, such as logistic regression,

on the device side. This limits the deployment of

mainstream neural networks that include feedback

processes (Yao, 2021). At the same time, to protect

user privacy, federated learning often employs

techniques such as differential privacy and

homomorphic encryption. While these technologies

provide privacy protection, they also increase

computational overhead and reduce training

efficiency. For example, the PrivacyFace framework

mentions that the Differential Privacy Clustering

Algorithm (DPLC), although capable of protecting

privacy, introduces additional computational

overhead, increasing the complexity of model

training. Therefore, the issue of high computational

resource demands is particularly prominent when

integrating facial recognition technology with

federated learning.

3.1.2 Communication Issues

Firstly, palm print recognition models typically use

deep learning methods, such as convolutional neural

networks (CNNs), which contain a large number of

parameters. In addition, the palm print image itself

has a large amount of data, especially in high-

resolution images. For example, a high-resolution

image of a palmprint may be several megabytes in

size, while a typical deep learning model may contain

millions of parameters. This large amount of data will

occupy a large amount of network bandwidth during

the transmission process, resulting in low

communication efficiency, and federated learning

requires frequent communication between the client

and the server, which will bring a huge burden to the

transmission network and greatly prolong the training

time. So the communication requirements of the

coordinator are high, as it often needs to wait for all

participants to deliver their intermediary data before

performing secure aggregation or other data

processing. One of the issues that hinders the

development of federated learning is how to improve

the capacity and quality of communication channels

(Konečný et al., 2017).

3.1.3 Security and Privacy Issues

Biometric data (such as fingerprints, faces, iris, palm

prints, etc.) is highly sensitive personal information.

Unlike traditional usernames and passwords,

biometric data is unique and unchangeable. Once

leaked, users will face serious privacy violations and

security risks. In federated learning, biometric data

needs to be transferred frequently between the client

and the server. However, most of the current research

in the relevant literature assumes that the participants

and servers of federated learning are secure and

trustworthy, but in practical applications, there is a

possibility that the user-sensitive model parameter

information obtained during the training process may

be exposed to the server or a third party, for example,

in an unstable network environment, data

transmission errors may occur, resulting in packet

loss or corruption. Attackers can exploit these errors

to obtain data through replay attacks, among other

things. During data transmission, attackers can also

disguise themselves as legitimate communication

nodes to intercept and tamper with data. This will

cause privacy leakage and reduce the security of

federated learning systems (Sun et al., 2022).

3.2 Future Prospects

Federated learning in open-world environments is a

new research domain. Despite notable research

advancements, some challenges need further

investigation, which is currently quite limited.

Initially, to mitigate the demand for computing

resources, two methodologies—edge cloud

collaborative computing and distributed computing

optimization—are proposed as effective strategies to

Advancements and Applications of Federated Learning in Biometric Recognition

529

address the high demand for computing resources and

the low communication efficiency encountered in the

integration of biometric technology and federated

learning. Edge-cloud collaborative computing

enables the offloading of complicated computational

activities to the cloud, alleviating the strain on edge

devices and optimizing the utilization of high-

performance computing resources in the cloud. By

optimizing distributed computing, one may

dynamically modify the computational tasks and

communication frequency of each device, equilibrate

the computational load among various devices, and

diminish communication overhead. The integration

of these two methodologies can markedly enhance the

computational efficiency of federated learning in

biometric applications.

The second is the need to improve communication

efficiency. This problem mainly stems from the large

amount of data in the recognition model of palmprint,

iris and other features, the difficulty of transmission

optimization, and the complex communication

coordination between different devices. In order to

solve this problem, the following strategies can be

adopted: asynchronous communication and data

deduplication technology can be used to improve

communication efficiency; Leverage edge computing

technology to reduce direct communication between

devices. Through these methods, the communication

efficiency of federated learning in biometric tasks can

be significantly improved.

Finally, in order to address the risk of data

breaches, a variety of measures can be taken, such as

data encryption, secure transmission protocols and

authentication etc. Through these measures, the data

security and privacy protection level of federated

learning in biometric tasks can be effectively

improved.

4 CONCLUSIONS

This article presents three categories of biometrics,

each associated with certain limitations. In order to

address the limitations of biometric recognition in

practice, it is possible to mitigate potential risks by

integrating it with federated learning. However,

federated learning continues to encounter challenges.

While it can safeguard personal privacy under

specific circumstances, challenges in data processing

and inefficiency must yet be resolved. In conclusion,

within the current framework prioritizing data

ownership and privacy safeguards, federated learning

possesses significant promise. However, several

technological hurdles and integration advancements

must be resolved. This review aims to offer support

and inspiration for research in federated learning-

based biometrics recognition.

REFERENCES

Aggarwal, D., Zhou, J., & Jain, A. K. 2021. Fedface:

Collaborative learning of face recognition model. In

2021 IEEE International Joint Conference on

Biometrics (IJCB) (pp. 1-8).

Chen, X., Jia, W., Li, S., et al. 2019. Palmprint recognition

method integrating global and local directional features

(in Chinese). Journal of Graphics, 40(4), 671-680.

Daugman, J. G. 1993. High confidence visual recognition

of persons by a test of statistical independence. IEEE

Transactions on Pattern Analysis and Machine

Intelligence, 15(11).

Daugman, J. 1998. Recognizing people by their iris

patterns. Information Security Technical Report, (1).

Gao, M. 2024. Research on face detection and recognition

method based on federated learning (in Chinese).

Doctoral dissertation, Changchun University of

Technology.

Huang, D., Jia, W., & Zhang, D. 2008. Palmprint

verification based on principal lines. Pattern

Recognition, 41(4), 1316-1328.

Jia, W., Zhu, Y., Liu, L., et al. 2009. Fast palmprint retrieval

using principal lines. In 2009 IEEE International

Conference on Systems, Man and Cybernetics (pp.

4118-4123). San Antonio: IEEE.

Konečný, J., McMahan, H. B., Yu, F. X., et al. 2017.

Federated learning: Strategies for improving

communication efficiency [EB/OL].

Li, C., Liu, F., & Zhang, Y. 2010. A principal palm-line

extraction method for palmprint images based on

diversity and contrast. In 2010 3rd International

Congress on Image and Signal Processing (pp. 1772-

1777). Yantai: IEEE.

Li, M. 2020. Research on the current situation and

development trend of fingerprint recognition

technology (in Chinese). Wireless Internet Technology,

17(01), 158-159.

Lian, F. 2023. Research on finger vein recognition based on

personalized federated learning (in Chinese). Doctoral

dissertation, South China University of Technology.

Meng, Q., Zhou, F., Ren, H., et al. 2022. Improving

federated learning face recognition via privacy-agnostic

clusters. arXiv preprint arXiv:2201.12467.

Sun, E., Zhang, H., He, R., et al. 2022. Mobile

communication resource management based on

federated learning: Methods, progress, and prospects

(in Chinese). Journal of Beijing University of

Technology, 48(07), 783-793.

Wang, C., Lu, Y., & Shen, J. 2024. Federated learning

fingerprint recognition algorithm for fairness (in

Chinese). Computer Science, 51(S1), 1014-1022.

Yao, J. J., & Ansari, N. 2021. Enhancing federated learning

in fog-aided IoT by CPU frequency and wireless power

DAML 2024 - International Conference on Data Analysis and Machine Learning

530

control. IEEE Internet of Things Journal, 8(5), 3438-

3445.

Zhang, Y., Liu, L., & Cheng, J. 2022. Training method and

device for face recognition model based on federated

learning (in Chinese) [Patent]. Beijing:

CN202111509411.X.

Advancements and Applications of Federated Learning in Biometric Recognition

531