The Advancements and Future Prospects of Federated

Learning-Based Methods for Biometrics

Yifan Zhang

Artificial Intelligence, Beijing Normal University-Hong Kong Baptist University United International College,

Zhuhai, China

Keywords: Federated Learning, Biometrics, Privacy Preservation.

Abstract: Biometrics has gained great attention due to its effectiveness in security applications recently. Although Deep

Learning (DL) has proven useful in biometrics, but it requires huge, high-quality datasets and raises privacy

issues. Federated Learning (FL) provides a solution, which trains models on clients and sending updates to

the global model without sharing sensitive information. This paper reviews recent advances in applying FL

for biometrics, mainly focusing on its use in face recognition, iris recognition, palmprint recognition, and

finger vein recognition. Some FL-based methods such as FedFace, FedGC, PrivacyFace, and other methods

are introduced, and their contributions are discussed in this paper. These methods have greatly contributed to

privacy preservation and model performance. Despite the great progress, there are still great challenges and

limitations in handling model interpretability and applicability due to the non-independent and identically

distributed (non-IID) data and model complexity. The future prospects include enhancing model

interpretability through techniques such as SHapley Additive exPlanations (SHAP) and Local Interpretable

Model-agnostic Explanations (LIME) and improving applicability by implementing transfer learning and

domain adaptation. This paper provides suggestions and references for future research and concludes that FL

provides a promising path forward for biometrics.

1 INTRODUCTION

Biometrics, the science and technology of analyzing

biological data, is under significant attention as a

rapidly evolving field. Because the conventional

information-saving methods is facing great

challenges as the technological innovation day by

day, using biometrics as the password is gaining

increasing popularity, which have the features of

uniqueness and non-transferability, and show great

effectiveness in device unlocking, identity

verification, and other practical applications (Jain &

Kumar, 2012). Biometrics is mainly divided into two

categories. The first category is physiological

characteristics, such as fingerprints, iris recognition,

facial recognition, and the second category is

behavioral characteristics, which include gait analysis

and voice recognition. Nowadays, researchers found

that the traditional method used in biometrics may not

have satisfactory predictive accuracy due to some

special scenarios; for instance, monozygotic twins

https://orcid.org/0009-0000-2520-9036

have a high degree of similarity in facial appearance

(Srinivas et al., 2011). Thus, researchers turned their

attention to Deep Learning (DL), for it has great

ability in feature extraction and prediction, which

match well with the requirements of the biometrics

field (Minaee et al., 2023).

In recent years, Artificial Intelligence (AI) such as

DL have been widely used in fields like biology,

chemistry, and especially biometrics. As shown by

Saeed et al. (Saeed et al., 2022), deep learning

methods based on Convolutional Neural Network

(CNN) model have achieved great progress on

classifying cross-sensor fingerprints. Similarly, the

use of Deep Learning techniques has proven effective

in face recognition and identification, as indicated by

the research of Teoh et al. (Teoh et al., 2021).

However, the high demand for biological data when

using deep learning requires considerable resources

and complicated data collection processes. Some

approaches have tried to solve the problem through

data sharing, but it may lead to leakage problems.

Zhang and Y.

The Advancements and Future Prospects of Federated Learning-Based Methods for Biometrics.

DOI: 10.5220/0013525300004619

In Proceedings of the 2nd International Conference on Data Analysis and Machine Learning (DAML 2024), pages 423-427

ISBN: 978-989-758-754-2

423

Therefore, this has led to increasing concerns about

the potential risk of revealing some sensitive user

information.

In this regard, researchers have considered the

need to implement Federated Learning (FL) to

biometrics to protect user privacy. FL allows the

client to update the model locally and only exchange

model parameters, offering a secure and collaborative

framework without sharing private and sensitive data

(Tomashenko et al., 2022). According to the work

form Cheng et al. (Cheng et al., 2022), a fingerprint-

based localization system has gained progress after

introducing a FL-based localization algorithm named

FedLoc-AC. Additionally, the work form Gupta et al.

on iris database performed in Federated Learning

with CNN also gained effective outcomes (Gupta et

al., 2022). In the same vein, FL has also been used in

behavioral characteristics such as fall detection and

automatic speech recognition (ASR), demonstrating

the adaptability of FL to privacy protection

requirements (Qi et al., 2023; Mammen, 2021). Since

FL can effectively train biometric models while

addressing the problem of user privacy protection,

and numerous works have been proposed recently, it

is vital to conduct a comprehensive review of the

application of FL to the field of biometrics.

The remainder of the paper is organized as follows.

In Section 2, this paper will summarize how to

combine FL with biometrics to train models and make

predictions under various situations. Then in Section

3, this paper will discuss some issues and limitations

that exist in recent works, in the meantime exploring

the future directions of Federated Learning in

biometrics. Finally, the whole paper will be

summarized in Section 4.

2 METHODS

2.1 Preliminaries of Federated

Learning

Federated Learning, as a decentralized machine

learning method, enables the clients to train a global

model without sharing their local data. The main idea

is that the central server only needs to aggregate these

updates from the client to improve the global model

to achieve preserving data privacy by using

distributed data sources.

The basic components of FL include clients, the

central server, and the communications between

them. Clients could be the node or the devices that

storing the data, and the gradients, weights, and

model changes are sent to the central server. A typical

FL workflow consists of three steps: local training,

model aggregation, and global model updates. The

whole process iterates from multiple communication

rounds until the central server’s global model

converges.

2.2 Federated Learning in Face

Recognition

2.2.1 FedFace

Bai et al. proposed a new federated learning

framework called FedFace to the problem of face

recognition (Bai et al., 2021). The framework

contains two new algorithms, Partially Federated

Momentum (PFM) and Federated Validation (FV).

PFM applies the estimated equivalent global

momentum locally to obtain approximate centralized

momentum to solve the client drift problem while

protecting training efficiency. The FV algorithm

dynamically searches for the best federated

aggregation weights by testing the aggregated model,

thus enhancing the ability of generalization. Through

experiments, FedFace has better privacy preservation

and performance compared to formal FL methods.

2.2.2 FedGC

Niu and Deng proposed a new federated learning

method called FedGC in the face recognition field

(Niu et al., 2022). This method introduces a SoftMax

regularizer and a gradient correction mechanism at

the base of FL infrastructure to enhance privacy

protection and model performance. The innovation of

FedGC is that it ensures each client has their own

private, fully connected layer, and by precisely

injecting cross-client gradient terms, it effectively

solves the privacy-preserving and model performance

decline issues that may occur in traditional FL

methods. Through experiments on popular datasets,

the effectiveness of the FedGC has been

demonstrated.

2.2.3 PrivacyFace

A new federated learning method called PrivacyFace

was proposed by the study carried out by Meng et al.

to enhance the privacy protection in facial recognition

(Meng et al., 2022). By combining the tradition FL

with Differential Privacy (DP) techniques, the

method improves the model performance.

PrivacyFace innovates an algorithm called

Differentially Private Local Clustering (DPLC) for

generating privacy-agnostic clusters of class centres.

This approach ensures that no specific information

DAML 2024 - International Conference on Data Analysis and Machine Learning

424

can be learned, no matter in what condition.

Additionally, the study used a consensus-aware facial

recognition loss function that can learn more

discriminative features. The framework is proven to

be differentially private by mathematics while

introducing a lightweight overhead and achieving

performance improvements.

2.3 Federated Learning in Iris

Recognition

2.3.1 FedAvg-based CNN

Gupta et al. (Gupta et al., 2022), in their paper, put

forward a method on iris recognition based on

federated learning to protect the user’s data privacy.

In this work, the researchers utilized the CNN model

to extract features and used Softmax as an activation

function to enhance the functionality of currently in

use frameworks. In addition, this paper discussed the

influence of the number of clients and the training

sessions on the model’s effectiveness. This approach

improved the accuracy of the iris recognition model

while avoiding the transmission of sensitive user data.

2.3.2 FedIris

Luo et al. introduced a new federated learning

framework called FedIris to improve the accuracy and

privacy protection of iris recognition (Luo et al.,

2022). In this framework, researchers proposed a Fed-

Triplet loss function based on iris templates instead

of typical averaging Empirical Risk Minimization

(ERM) to enhance the interpretability. Besides, by

adjusting the Wasserstein Distance to reweighting

aggregation and incorporating it into Fed-Triplet loss,

the high distribution shifts between the clients by the

negative transfer will be mitigated. Experiments on

multiple public iris data verify that FedIris surpasses

model-sharing FL methods.

2.4 Federated Learning in Palmprint

2.4.1 PSFed-Palm

The work from Yang et al. proposed a novel physics-

driven spectrum-consistent federated learning

method named PSFed-Palm for palmprint

verification (Yang et al., 2024). The method uses the

physical properties of different wavelength spectrums

and partitions the client into two groups: short-

wavelength and long-wavelength groups based on

their spectrum types, then designs the corresponding

anchor models and constrains the optimization

directions. Besides, the work designed a spectrum-

consistent loss function to ensure the consistency

between different spectral templates. In this situation,

the model can prevent model drift and protect data

privacy without sharing local data, showcasing its

vast potential for palmprint verification.

2.5 Federated Learning in Finger Vein

2.5.1 FedFV

Lian et al. proposed a federated learning framework

called FedFV for finger vein certification (Lian et al.,

2023). Through the use of FedFV, each client can

acquire information from all members of the

federation without resulting in the leakage of

templates. Furthermore, the work also proposes a

personalized federated aggregation algorithm called

FedWPR in order to solve the issue of non-IID data

brought on by client diversity, therefore, to obtain the

best performance for each client. In addition, FedFV

performed well on different scenarios, as the first

personalized federal finger vein authentication

framework, will also serve as a point of reference for

upcoming studies on biometric privacy protection.

2.5.2 fvPAD

Mu et al. proposed a novel federated learning method

for finger vein presentation attack detection (fvPAD)

(Mu et al., 2024). The work first applied FL in fvPAD

by considering the data volume and computational

power among different clients and divided the

traditional FL clients into two categories: institutional

clients and terminal clients. For institutional clients,

the work designed a triplet training mode combined

with FL to enhance the model's generalization.

Meanwhile, the study introduced an unsupervised

learning module for terminal clients to enable the

fvPAD model to adopt local data distribution and

optimize itself while using unlabelled data. Through

extensive experiments, the results showed the

advantage in accuracy and robustness, especially

when dealing with data from diverse clients.

3 DISCUSSIONS

3.1 Challenges and Limitations

3.1.1 Interpretability

Despite the significant contributions of FL to address

biometric issues, there also raise questions about the

The Advancements and Future Prospects of Federated Learning-Based Methods for Biometrics

425

model’s interpretability. Because FL is trained on

each client and the training process is non-

transparent, this will increase the complexity of

comprehending the model. Besides, the data held by

the client are difficult to be Independent Identically

Distribution (IID), therefore leading to data

heterogeneity, for which the model needs to be

generalized over different data distributions; this also

affects the interpretability of the model. In addition,

FL, through the exchange of the model update to

protect privacy instead of using raw biometric data,

may still encapsulate sensitive information.

Consequently, providing interpretability of the model

without any divulge is particularly challenging.

3.1.2 Applicability

In addition to the challenges about the

interpretability, the problem of applicability

associated with FL also poses a great challenge. Due

to the non-Independent and Identically Distributed

(non-IID) characteristics, such as label distribution

skew, feature distribution skew, and data quantity

skew, the model may be impeded to converge under

the condition that some clients are significantly

different in distribution from other clients. Thus, the

model of the client may overfit to the local data to

adapt to the data distribution, which can conflict with

the objectives of the global model. Therefore,

handling non-IID data in federated learning scenarios

is significant to duel with applicability issues.

3.2 Future prospects

To mitigate these issues, focusing on enhancing

model interpretability and using transfer learning and

domain adaptation are key objectives for future

research.

Introducing some existing instruments is efficient

to enhance model interpretability. For instance,

Expert System (ES) can help with comprehending

how biometric data is used to train and predict the FL

model. Based on the knowledge from expert systems,

constructing an explainable model becomes more

feasible. Additionally, implementing a popular

interpretive model SHapley Additive exPlanations

(SHAP) into FL in biometrics can help to understand

how each feature contributes to the predictions made

by the model, thereby enhancing the model's

interpretability. Moreover, using Local Interpretable

Model-agnostic Explanations (LIME) can be another

option. LIME can create interpretable models locally

based on specific prediction instances, hence

providing more intuitive explanations for the

predictions made by federated learning models in

biometrics. In addition, some advanced

interpretability methods can be also considered. For

example, Qiu et al. proposed a dense registration-

based approach for improving the interpretability

during the fingerprint matching process (Qiu, 2024).

Yin et al. focus on enhancing the interpretability of

face recognition by proposing a novel approach that

incorporates a spatial activation diversity loss (Yin,

2019). This method is designed to facilitate the

learning of more structured and interpretable face

representations.

The utilization of Transfer Learning (TL) and

Domain Adaptation (DA) can also facilitate future

research at FL in biometrics. TL allows models to

leverage knowledge acquired from one dataset to

enhance performance on another dataset. Because the

biometric data are rare and expensive, TL shows its

particular effort in biometrics when there are limited

labelled data. Meanwhile, biometric features tend to

have different distributions in different cultures and

regions captured by different devices. Owing to

Federated Learning based biometrics being an

interdisciplinary field, researchers can improve the

model’s applicability across various devices and

environments through domain adaptation.

The future improvement of federated learning in

biometrics tends to focus on how to improve the

applicability and interpretability without sacrificing

user privacy. By combining interpretability

techniques, transfer learning, and domain adaptation,

advancements will be gained in pursuit of this goal.

4 CONCLUSION

This paper provided a comprehensively review and

summarization about recent advances in biometrics

using Federated Learning. This paper has introduced

several FL-based methods such as FedFace, FedGC,

and PrivacyFace for face recognition and other

methods for various biometric subfields. The

integration of FL has led to significant advancements,

but it also presents new challenges that must be

navigated. The future prospects of FL in biometrics

are full of opportunities, some potential research

directions include mitigating non-IID data

challenges, enhancing model interpretability. As a

rapidly evolving field, the federated learning

paradigm's application in biometrics still have great

gaps to fulfill, but it has a promising future which user

biometric privacy is preserved and data utilization is

maximized. This paper aims to provide valuable

insights and reference for researchers to continue

pushing the boundaries in this field.

DAML 2024 - International Conference on Data Analysis and Machine Learning

426

REFERENCES

Bai, F., Wu, J., Shen, P., Li, S., & Zhou, S. 2021. Federated

face recognition. arXiv preprint arXiv:2105.02501.

Cheng, X., Ma, C., Li, J., Song, H., Shu, F., & Wang, J.

2022. Federated learning-based localization with

heterogeneous fingerprint database. IEEE Wireless

Communications Letters, 11(7), 1364-1368.

Gupta, H., Rajput, T. K., Vyas, R., Vyas, O. P., & Puliafito,

A. 2022. Biometric iris identifier recognition with

privacy preserving phenomenon: A federated learning

approach. In International Conference on Neural

Information Processing (pp. 493-504). Singapore:

Springer Nature Singapore.

Jain, A. K., & Kumar, A. 2012. Biometric recognition: an

overview. Second generation biometrics: The ethical,

legal and social context, 49-79.

Lian, F. Z., Huang, J. D., Liu, J. X., Chen, G., Zhao, J. H.,

& Kang, W. X. 2023. FedFV: A personalized federated

learning framework for finger vein authentication.

Machine Intelligence Research, 20(5), 683-696.

Luo, Z., Wang, Y., Wang, Z., Sun, Z., & Tan, T. 2022.

Fediris: Towards more accurate and privacy-preserving

iris recognition via federated template communication.

In Proceedings of the IEEE/CVF Conference on

Computer Vision and Pattern Recognition (pp. 3357-

3366).

Mammen, P. M. 2021. Federated learning: Opportunities

and challenges. arXiv preprint arXiv:2101.05428.

Meng, Q., Zhou, F., Ren, H., Feng, T., Liu, G., & Lin, Y.

2022. Improving federated learning face recognition via

privacy-agnostic clusters. arXiv preprint

arXiv:2201.12467.

Minaee, S., Abdolrashidi, A., Su, H., Bennamoun, M., &

Zhang, D. 2023. Biometrics recognition using deep

learning: A survey. Artificial Intelligence Review,

56(8), 8647-8695.

Mu, H., Guo, J., Liu, X., Han, C., & Sun, L. 2024. Federated

finger vein presentation attack detection for various

clients. IET Computer Vision.

Niu, Y., & Deng, W. 2022. Federated learning for face

recognition with gradient correction. In Proceedings of

the AAAI Conference on Artificial Intelligence (Vol.

36, No. 2, pp. 1999-2007).

Qi, P., Chiaro, D., & Piccialli, F. 2023. FL-FD: Federated

learning-based fall detection with multimodal data

fusion. Information fusion, 99, 101890.

Qiu, Y., Chen, H., Dong, X., Lin, Z., Liao, I. Y., Tistarelli,

M., & Jin, Z. 2024. Ifvit: Interpretable fixed-length

representation for fingerprint matching via vision

transformer. arXiv preprint arXiv:2404.08237.

Saeed, F., Hussain, M., & Aboalsamh, H. A. 2022.

Automatic fingerprint classification using deep learning

technology (DeepFKTNet). Mathematics, 10(8), 1285.

Srinivas, N., Aggarwal, G., Flynn, P. J., & Bruegge, R. W.

V. 2011. Facial marks as biometric signatures to

distinguish between identical twins. In CVPR 2011

WORKSHOPS (pp. 106-113). IEEE.

Teoh, K. H., Ismail, R. C., Naziri, S. Z. M., Hussin, R., Isa,

M. N. M., & Basir, M. S. S. M. 2021. Face recognition

and identification using deep learning approach. In

Journal of Physics: Conference Series (Vol. 1755, No.

1, p. 012006). IOP Publishing.

Tomashenko, N., Mdhaffar, S., Tommasi, M., Estève, Y.,

& Bonastre, J. F. 2022. Privacy attacks for automatic

speech recognition acoustic models in a federated

learning framework. In ICASSP 2022-2022 IEEE

International Conference on Acoustics, Speech and

Signal Processing (ICASSP) (pp. 6972-6976). IEEE.

Yang, Z., Teoh, A. B. J., Zhang, B., Leng, L., & Zhang, Y.

2024. Physics-Driven Spectrum-Consistent Federated

Learning for Palmprint Verification. International

Journal of Computer Vision, 1-16.

Yin, B., Tran, L., Li, H., Shen, X., & Liu, X. 2019. Towards

interpretable face recognition. In Proceedings of the

IEEE/CVF International Conference on Computer

Vision (pp. 9348-9357).

The Advancements and Future Prospects of Federated Learning-Based Methods for Biometrics

427