FedBID and FedDocs: A Dataset and System for Federated Document
Analysis
Daniel Perazzo
1,†
, Thiago de Souza
1,†
, Pietro Masur
1
, Eduardo de Amorim
1
, Pedro de Oliveira
1
,
Kelvin Cunha
1
, Lucas Maggi
1
, Francisco Sim
˜
oes
1,2
, Veronica Teichrieb
1
and Lucas Kirsten
3
1
Voxar Labs, Centro de Inform
´
atica, Universidade Federal de Pernambuco, Recife/PE, Brazil
2
Visual Computing Lab, Departamento de Computac¸
˜
ao, Universidade Federal Rural de Pernambuco, Recife/PE, Brazil
3
HP Inc., Porto Alegre/RS, Brazil
Keywords:
Federated Learning, Document Analysis, Privacy, Dataset.
Abstract:
Data privacy has recently become one of the main concerns for society and machine learning researchers. The
question of privacy led to research in privacy-aware machine learning and, amongst many other techniques,
one solution gaining ground is federated learning. In this machine learning paradigm, data does not leave the
user’s device, with training happening on it and aggregated in a remote server. In this work, we present, to our
knowledge, the first federated dataset for document classification: FedBID. To demonstrate how this dataset can
be used for evaluating different techniques, we also developed a system, FedDocs, for federated learning for
document classification. We demonstrate the characteristics of our federated dataset, along with different types
of distributions possible to be created with our dataset. Finally, we analyze our system, FedDocs, in our dataset,
FedBID, in multiple different scenarios. We analyze a federated setting with balanced categories, a federated
setting with unbalanced classes, and, finally, simulating a siloed federated training. We demonstrate that FedBID
can be used to analyze a federated learning algorithm. Finally, we hope the FedBID dataset allows more research
in federated document classification. The dataset is available in https://github.com/voxarlabs/FedBID.
1 INTRODUCTION
The fast development of learning-based techniques,
especially Deep Learning (DL), inspired new
applications in different areas of computing (Xu et al.,
2019). These applications became more accurate and
efficient through increasingly robust learning neural
networks and access to more data. For example,
numerous applications on computer vision, using deep
learning techniques, emerged in recent years in fields
such as robotics, autonomous driving, and assistive
technologies (S
¨
underhauf et al., 2018).
Nevertheless, for distributed training with edge
devices, deep learning has traditionally followed a
centralized training approach, in which third-party
servers were responsible for the entire training
process and where all data is sent and stored in the
server (Huang et al., 2017). In this type of training,
the user’s data needs to be sent, stored, and processed
on servers to train the neural network, which needs
to be retrained continuously over time (Mayr et al.,
†
Denotes equal contribution
2019). However, this type of centralized training is
highly susceptible to security breaches due to all data
being stored in a single server.
In this scenario, Federated Learning (FL) is a
new paradigm to distribute neural network training
with data privacy as a main requirement (McMahan
et al., 2017). In this setting, part or all of the training
process is addressed in remote devices, minimizing the
exchange of sensitive data from users since customers’
data is never sent to the network or accessed by an
external agent directly. The models are updated over
time, adding new information that reflects the data
distribution for each user.
Over the last few years, different strategies and
architectures have been proposed in FL for different
areas (Liu et al., 2020; Li et al., 2019; Zhu et al.,
2021). However, consolidated benchmarks and
datasets are lacking in the FL literature to validate
applications (Caldas et al., 2018). In most cases,
datasets are either not publicly available, overly simple,
or do not have adequate distribution to assess FL
challenges (Caldas et al., 2018).
In this context, we present a new dataset that
Perazzo, D., de Souza, T., Masur, P., de Amorim, E., de Oliveira, P., Cunha, K., Maggi, L., Simões, F., Teichrieb, V. and Kirsten, L.
FedBID and FedDocs: A Dataset and System for Federated Document Analysis.
DOI: 10.5220/0011658700003417
In Proceedings of the 18th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2023) - Volume 5: VISAPP, pages
551-558
ISBN: 978-989-758-634-7; ISSN: 2184-4321
Copyright
c
 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)
551

contains a well-defined non-IID distribution. The
BID (de S
´
a Soares et al., 2020) is a dataset for identity
document classification, containing annotations for
different types of Brazilian documents. Through
analyzing BID, we could identify data patterns and
classes that we can map to existing challenges
in Federated Learning. In this way, we propose
FedBID, a new dataset in federated document
classification containing 21, 600 examples annotated
from an accurate non-IID distribution, to evaluate
new challenges not mapped by previous datasets.
Among these challenges, FedBID can consider test
scenarios with a severe imbalance of examples in
different users, with variations related to classes,
number of samples, image quality, and annotation
reliability. Furthermore, FedBID was designed to
handle various tasks such as labeling the document
type, orientation, source organization, and document
owner. We also constructed a system, FedDocs, for
federated document classification and evaluated the
accuracy of FedDocs in our dataset.
As the main contributions, we highlight the
following:
• FedBID, the first federated document classification
dataset (Section 4);
• FedDocs, a system for federated learning for
document classification (Section 3);
• Finally, we evaluate our proposed system on
FedBID, with various configurations for federated
learning. (Section 5).
2 RELATED WORKS
Federated learning is a new paradigm for decentralized
learning (McMahan et al., 2017) and a relatively
new research field, with many challenges and open
problems still present in the research community, as
shown in recent surveys (Kairouz et al., 2021; Li et al.,
2020). Below we list research related to our work.
Federated Learning Datasets. Due to its
decentralized training procedure, federated learning
datasets differ from traditional machine learning
datasets, requiring data partitions for various devices.
Many researchers have proposed and released datasets
and evaluation benchmarks for federated learning.
An essential factor for these datasets is for them to
mimic non-IID distributions since they frequently
occur in the real world and are still a challenge for
federated learning algorithms (Shoham et al., 2019). A
longstanding issue with federated learning datasets is
that creating a federated non-IID dataset by artificially
partitioning datasets commonly used for machine
learning (e.g., CIFAR, MNIST) can create datasets
that do not follow a realistic non-IID distribution.
Due to this issue, (Caldas et al., 2018) created a
benchmark comprised of multiple federated datasets
and metrics. For example, their federated version
of MNIST (FEMNIST) is partitioned according
to the writer of the digit. In contrast, their dataset
Sentiment140, for sentiment analysis, is partitioned
based on the Twitter user. Furthermore, (He et al.,
2020) introduced FedML, a framework for federated
learning that includes default datasets for users to get
started. Finally, (Koh et al., 2021; Luo et al., 2019)
created datasets focused on real-world and in-the-wild
data.
Federated Learning Applications. Numerous
applications were proposed for algorithms using
federated learning. For example, in NLP, Google
has used federated learning to improve query
suggestions in the “Google Keyboard” while
preserving user privacy (Yang et al., 2018). Other
applications of federated learning were also in emoji
prediction (Ramaswamy et al., 2019), also by Google,
and speech recognition (Paulik et al., 2021), by
Apple. In computer vision, FedVision (Liu et al.,
2020) introduced a platform for object detection
using federated learning. Furthermore, one field
that has been receiving attention in federated
learning is healthcare, where patient data is very
sensitive (Antunes et al., 2022). Finally, with the
expansion of edge devices and the internet of things,
federated learning is increasingly receiving attention
due to possible applications for training on edge
devices (Kontar et al., 2021)
Document Analysis. Document analysis is a
longstanding problem in computer vision with more
than two decades of research (Liu et al., 2021).
Although in the past decades, researchers focused their
work on traditional computer vision techniques, such
as image processing and pattern matching, to perform
their analysis (Love et al., 2013), more recently,
machine learning approaches have shown promising
results (Li et al., 2021). For example, (Kang et al.,
2014) interpreted document classification as an image
classification problem and used Convolutional Neural
Networks (CNNs) to perform the task. Researchers
aiming to stimulate progress in the area created
numerous new datasets, competitions, and benchmarks
for document analysis using data from many different
countries (Burie et al., 2015; Bulatovich et al.,
2022a; Chernyshova et al., 2021; Bulatovich et al.,
2022b; Polevoy et al., 2022). In this context,
VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications
552

(de S
´
a Soares et al., 2020) created a document
classification based on Brazilian personal documents.
Document identification is a fascinating case study
for federated learning since privacy is a significant
concern among users because malicious users can use
the leaked information associated with identification
details to perpetrate financial fraud, false identity, and
many other crimes.
3 FedDocs
In this section, we explain our FedDocs system. To
do so, we will detail more about its architecture, the
process in which the clients are trained, and how the
server aggregates the weights.
3.1 Architecture of the System
Our application simulates a remote Federated Learning
(FL) architecture in different devices and networks.
The module comprises two applications: a desktop
client and a server app. Both are written in Python
and ran over the docker virtualization engine. The
docker image contains the Flower library (Beutel et al.,
2020), a general FL framework. Flower provides tools,
data structures, and protocols to perform federated
learning training across different devices, using the
gRPC (Google Remote Procedure Call) framework
(Marculescu, 2015).
Both applications (Client and Server modules)
are the primary tools of our architecture and can be
used in a real scenario to deploy the applications
on different devices. Moreover, the architecture
provides tools to adapt various applications or datasets
for supervised image classification quickly. In our
case, more specifically, dealing with image document
classification.
3.2 Client Training
The desktop client is responsible for locally training
the received model from the server, using the available
data, and sending back the updated parameters to
the server. We perform this training similarly to the
standard centralized setting, where each local data
sample contributes to the learning progress in every
epoch. The process finishes when it achieves a stop
criteria, such as the maximum number of local epochs
or a minimum loss value. In this paper, we use the
number of local epochs processed on each device with
their local dataset as the stop criteria.
After the training is concluded in all clients, the
trained model weights are sent back from the clients
to the server for the resulting aggregated model. Then
the server joins the knowledge and information of each
client neural network model into one global neural
network model.
3.3 Server Aggregation
The server coordinates the training process, aggregates
the client parameters (i.e., weights) in each round, and
retrieves the training metrics (i.e., accuracy, number
of samples). We compute the total accuracy by taking
the mean of the accuracy for each device on their local
test set. To implement this task, we use the strategy
abstraction in Flower, where each strategy provides
instructions to train and evaluate models on clients,
and perform the aggregationon the server.
As aforementioned, the connected clients send the
updated weights to the server for the aggregation phase.
Hence, no data leaves the local device, and only the
model parameters go through the network.
In each round (i.e., training iteration), the server
aggregates the local client model’s weights through an
aggregation method, in our case FedAvg (McMahan
et al., 2017). FedAvg makes a weighted average of the
local models and generates a new aggregated global
model. Then, this new aggregated global model is sent
back to the clients for evaluation. Finally, new clients
are selected for the next training cycle. The process
repeats until the model converges or achieves a stop
criteria.
4 FedBID
In this section, we describe our new dataset FedBID,
based on the BID Dataset (de S
´
a Soares et al., 2020).
4.1 Summary of BID Dataset
BID Dataset (Brazilian Identity Document Dataset)
is a dataset composed of images of the Brazilian
Driver’s License (a.k.a. CNH), Natural Person registry
document (a.k.a. CPF), and the Brazilian identification
document (a.k.a. RG) that aims to help researchers
with numerous challenges of computer vision for
document automated processing, such as classification
and segmentation, etc. We used this dataset since it
was generated using a process that anonymizes the
original publicly available data, creating a dataset that
complies with Brazilian data privacy laws (LGPD,
2018). Based on the original BID dataset, we manually
created a new set of labels and a distribution of
documents per various clients, as we proceed to detail
shortly.
FedBID and FedDocs: A Dataset and System for Federated Document Analysis
553

The setup of the original BID dataset allows a
classification problem between the previously stated
document classes (CPF, CNH, and RG) and their
respective back and front counterparts. In this
manner, we have the following six classes: CPF Front,
CPF Back, CNH Front, CNH Back, RG Front, and
RG Back.
4.2 Distribution and Methodology
BID was initially conceived for the utilization of
segmentation ML systems. Each BID image has
a corresponding label with information from the
image segmentation. Furthermore, images of different
documents are categorized into different folders. This
setting, although sufficient for traditional classification
systems, lacks some information that may be relevant,
such as orientation, which can be important for some
applications, such as OCR systems.
Using the BID as a baseline, we further build
another system of classes. Since determining the
orientation of a document is essential for many OCR
systems to extract the text correctly, we added this type
of class to the BID dataset. To do this, we manually
annotate each document image with its orientation
( 0
◦
, 90
◦
, 180
◦
, 270
◦
) using an interface from (Goecks
et al., 2021). Since we have four different orientations
with six classes of documents, we created a set of 24
categories by a cartesian product.
Furthermore, some of the documents on the BID
dataset contain information regarding the emission
issuer. To condense this information into a single label
file, we have built a script for automatically retrieving
this information from the segmentation labels. It works
by, for each image, matching its label file context with
a list containing all the document issuers in Brazil.
We can use the document issuer labels to simulate a
cross-silo federated data distribution. For example, in
Figure 2, we detail the distribution of the document
issuers found on the dataset. Details about the FedBID
dataset are shown in Table 1 and Figure 1. Separating
these issuers is interesting since, with this information,
we can perform a simulation of a siloed federated
learning setting.
Table 1: FedBID dataset description.
Split Samples Average Per Class
All Data 21600 3600 ± 0
Train 15120 2520 ± 12.5
Test 6480 1080 ± 12.5
Figure 1: The distribution for the 24 classes. The
FedBID distribution is unbalanced and non-IID. The classes
in this histogram are labeled in the following manner:
“(document class) (Front/Back) (angle)”.
Figure 2: Details on the nine different issuers present
on FedBID. Those documents that do not provide this
information are displayed in the unknown column.
4.3 Available Distributions
To verify different settings for federated algorithms,
we included different distributions of our data into
different distributions to evaluate many aspects
of federated training, such as initializing with a
pre-trained model on some categories. In this setting,
we use all 24 categories on each device, naming it a
balanced dataset. To evaluate algorithms, we built
three scenarios with 100, 500, and 1000 clients for
each distribution below:
• Full Dataset: All samples from the entire dataset
are partitioned almost equally between the devices,
without repetition;
• 50% of train data for centralized training: We
separate 50% from the entire dataset to pre-train a
centralized model.
We also wanted to observe the effect of different
categories on each device. To do so, we change the
range of the number of categories in each device
for these distributions, randomly removing some
classes from each client and partitioning the training
dataset into 100 clients. We name this distribution a
unbalanced dataset.
VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications
554

5 EXPERIMENTS
In this section, we propose a series of experiments
that aim to demonstrate how our dataset can be used to
benchmark a federated document classification system,
in our case, the FedDocs system.
5.1 Training Setup
We use our system, FedDocs, with the FedAvg
aggregator for all experiments and the same training
local configuration. The chosen model architecture is
the EfficientNet B0 (Tan and Le, 2019). We resize the
documents’ image inputs to a resolution of 224x224
RGB pixels. The number of local epochs is set
to E = 1 , and the batch size is B = 16 . We use
the Adam (Kingma and Ba, 2014) optimizer with
learning-rate lr = 0.0005.
5.2 Results
In this section, we present the results of our
experiments with our federated learning tests. All
reported results are related to our test set.
5.2.1 Federated Training with a Balanced
Dataset
In Figure 3, we present the training results of our
solution in the BID dataset with no pre-trained model
for different numbers of clients and use the balanced
dataset distributed for each device. We evaluate the
model after training it for one epoch.
Figure 3: Federated results. Each curve represents a training
with a different number of total clients: (blue): 100 clients;
(red): 500 clients; (yellow): 1000 clients.
As presented in Figure 4, using a model pre-trained
on 50%, the results are similar even if we increase the
number of clients. Although, as we note, the federated
training does not seem to help gain extra accuracy
points, the training course appears to decrease the
accuracy of the global model slightly.
Figure 4: Federated training results starting from a
centralized model pre-trained with 50% of training data.
Each curve represents a training with a different number of
total clients: (blue): 100 clients; (red): 500 clients; (yellow):
1000 clients.
In this case, we can see some challenges in
our dataset that researchers could address with new
federated techniques, for example, a catastrophic
forgetting (Kirkpatrick et al., 2017) of a pre-trained
model as seen in the work of (Shoham et al.,
2019). Thus, FedBID could be considered a dataset
contribution for future Federated Learning research.
5.2.2 Federated Training with an Unbalanced
Dataset
Figure 5: Unbalanced federated results without pre-trained.
(blue): The classes vary between 18 and 24. (red): The
classes vary between 12 and 24
In Figures 5 and 6, we present the results of training
our model in the unbalanced dataset, with a varying
range of possible categories on each device. We
can see similar behavior to federated training in the
balanced setting. In the tests without the pre-trained
centralized models using all categories, it is noticeable
that the convergence is slower since the models only
reach more than 50% of accuracy near round 200,
while for the unbalanced setting, we obtained it around
round 100. Faster convergence is crucial since it saves
clients’ data transfer and battery consumption. The
difference between varying the range of categories
in each device is small, even with the reduction
FedBID and FedDocs: A Dataset and System for Federated Document Analysis
555

Figure 6: Unbalanced federated training results starting from
a centralized model pre-trained with 50% of training data.
(blue): The classes vary between 18 and 24. (red): The
classes vary between 12 and 24
of categories between them. However, increasing
this unbalance in categories between the clients can
make learning processing more challenging. These
scenarios can be helpful to test new methods to deal
with the convergence in unbalanced configurations,
for example, using the FedProx technique (Li et al.,
2020), since the local domain diverges from the global
domain.
In the tests with the pre-trained model (Figure 4),
we can observe a similar behavior as the experiments
with all categories, with the results showing a
tendency to the neural network models “forgetting”
the centralized knowledge, and, as a consequence,
don’t learning between the rounds. Therefore, we
can explore this scenario’s previous challenges in the
unbalanced scenario.
5.2.3 Federated Training with Silloed Data
In this test, our architecture uses the nine natural
partitions from FedBID according to the document
issuer, presented in Figure 2, where each one has
only one type of document, but with the document
style according to its owner, except the unknown
partition which has classes from many unknown users
and different document types. This setting is near
to a real scenario of federated training where the
data is non-IID. Our architecture results showed this
distribution’s impact on model convergence.
As shown in Figure 7, the model convergence does
not increase gradually as in the IID scenarios. On
the contrary, the curve appeared much fuzzier with
jumps between 6% to 51% of accuracy without a
clear indication of when the model will converge.
This situation is caused due to the difference again
between the local domains and the global, which, now,
for this dataset, is the most diverse due to the data
heterogeneity level.
On the other hand, in Figure 8, we tested the
same number of partitions from the last experiment
Figure 7: Non-idd federated training results using siloed
partition where each contains data from one exclusive issuer.
Figure 8: Federated training results using nine siloed
partitions containing balanced IID data from all issuers.
now with all data distributed in a balanced form, we
can see a now expected smooth convergence with a
gradually increasing accuracy. This result highlights
the challenge of dealing with a non-IID distribution
since this IID scenario is not possible in a real case due
to privacy, opening a new set of federated challenges
for document categorization.
6 CONCLUSION AND FUTURE
WORKS
Our experiments show that our new FedBID dataset
can be a benchmark for federated learning in Non-IID
tasks. To perform the tests, we created an application
named FedDocs to perform federated document
classification. Our new federated dataset gives the
first federated dataset for document classification. For
the federated learning community, we also provide a
dataset with an attractive property, a natural federated
distribution based on issuers. As a result, researchers
could explore cross-silo labeling in future works. In
future work, we intend to construct federated learning
algorithms for non-IID distributions exploring our
dataset for future work.
VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications
556

ACKNOWLEDGMENT
This paper was achieved in cooperation with HP
Inc. R&D Brazil, using incentives of the Brazilian
Informatics Law (n°. 8.2.48 of 1991). The authors
would like to thank Thomas da Silva Paula for being
the project conceiver and Eduardo Chagas for the
management support during this work.
REFERENCES
Antunes, R. S., Andr
´
e da Costa, C., K
¨
uderle, A., Yari,
I. A., and Eskofier, B. (2022). Federated learning
for healthcare: Systematic review and architecture
proposal. ACM Transactions on Intelligent Systems
and Technology (TIST).
Beutel, D. J., Topal, T., Mathur, A., Qiu, X., Parcollet, T.,
de Gusm
˜
ao, P. P., and Lane, N. D. (2020). Flower: A
friendly federated learning research framework. arXiv
preprint arXiv:2007.14390.
Bulatovich, B. K., Vladimirovna, E. E., Vyacheslavovich,
T. D., Sergeevna, S. N., Sergeevna, C. Y., Zuheng, M.,
Jean-Christophe, B., and Muzzamil, L. M. (2022a).
Midv-2020: A comprehensive benchmark dataset
for identity document analysis. Computer Optics,
46(2):252–270.
Bulatovich, B. K., Vladimirovna, E. E., Vyacheslavovich,
T. D., Sergeevna, S. N., Sergeevna, C. Y., Zuheng, M.,
Jean-Christophe, B., and Muzzamil, L. M. (2022b).
Midv-2020: A comprehensive benchmark dataset for
identity document analysis. 46(2):252–270.
Burie, J.-C., Chazalon, J., Coustaty, M., Eskenazi, S.,
Luqman, M. M., Mehri, M., Nayef, N., Ogier,
J.-M., Prum, S., and Rusi
˜
nol, M. (2015). Icdar2015
competition on smartphone document capture and ocr
(smartdoc). In 2015 13th International Conference on
Document Analysis and Recognition (ICDAR), pages
1161–1165. IEEE.
Caldas, S., Duddu, S. M. K., Wu, P., Li, T., Kone
ˇ
cn
`
y, J.,
McMahan, H. B., Smith, V., and Talwalkar, A. (2018).
Leaf: A benchmark for federated settings. arXiv
preprint arXiv:1812.01097.
Chernyshova, Y., Emelianova, E., Sheshkus, A., and
Arlazarov, V. V. (2021). Midv-lait: a challenging
dataset for recognition of ids with perso-arabic, thai,
and indian scripts. In International Conference on
Document Analysis and Recognition, pages 258–272.
Springer.
de S
´
a Soares, A., das Neves Junior, R. B., and Bezerra,
B. L. D. (2020). Bid dataset: a challenge dataset for
document processing tasks. In Anais Estendidos do
XXXIII Conference on Graphics, Patterns and Images,
pages 143–146. SBC.
Goecks, V. G., Waytowich, N., Watkins, D., and Prakash, B.
(2021). Combining learning from human feedback and
knowledge engineering to solve hierarchical tasks in
minecraft. arXiv preprint arXiv:2112.03482.
He, C., Li, S., So, J., Zeng, X., Zhang, M., Wang, H., Wang,
X., Vepakomma, P., Singh, A., Qiu, H., et al. (2020).
Fedml: A research library and benchmark for federated
machine learning. arXiv preprint arXiv:2007.13518.
Huang, Y., Ma, X., Fan, X., Liu, J., and Gong, W. (2017).
When deep learning meets edge computing. In
2017 IEEE 25th international conference on network
protocols (ICNP), pages 1–2. IEEE.
Kairouz, P., McMahan, H. B., Avent, B., Bellet, A.,
Bennis, M., Bhagoji, A. N., Bonawitz, K., Charles, Z.,
Cormode, G., Cummings, R., et al. (2021). Advances
and open problems in federated learning. Foundations
and Trends® in Machine Learning, 14(1–2):1–210.
Kang, L., Kumar, J., Ye, P., Li, Y., and Doermann, D. (2014).
Convolutional neural networks for document image
classification. In 2014 22nd International Conference
on Pattern Recognition, pages 3168–3172. IEEE.
Kingma, D. P. and Ba, J. (2014). Adam: A
method for stochastic optimization. arXiv preprint
arXiv:1412.6980.
Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J.,
Desjardins, G., Rusu, A. A., Milan, K., Quan, J.,
Ramalho, T., Grabska-Barwinska, A., et al. (2017).
Overcoming catastrophic forgetting in neural networks.
Proceedings of the national academy of sciences,
114(13):3521–3526.
Koh, P. W., Sagawa, S., Marklund, H., Xie, S. M., Zhang,
M., Balsubramani, A., Hu, W., Yasunaga, M., Phillips,
R. L., Gao, I., et al. (2021). Wilds: A benchmark
of in-the-wild distribution shifts. In International
Conference on Machine Learning.
Kontar, R., Shi, N., Yue, X., Chung, S., Byon, E.,
Chowdhury, M., Jin, J., Kontar, W., Masoud, N.,
Nouiehed, M., et al. (2021). The internet of federated
things (ioft). IEEE Access.
LGPD (2018). General law for personal data protection
(lgpd). http://www.planalto.gov.br/ccivil 03/
ato2015-2018/2018/lei/L13709.htm. Accessed:
2022-09-19.
Li, Q., Wen, Z., Wu, Z., Hu, S., Wang, N., Li, Y., Liu, X.,
and 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.
Li, T., Sahu, A. K., Talwalkar, A., and Smith, V. (2020).
Federated learning: Challenges, methods, and future
directions. IEEE Signal Processing Magazine.
Li, W., Milletar
`
ı, F., Xu, D., Rieke, N., Hancox, J., Zhu, W.,
Baust, M., Cheng, Y., Ourselin, S., Cardoso, M. J., et al.
(2019). Privacy-preserving federated brain tumour
segmentation. In International workshop on machine
learning in medical imaging.
Liu, L., Wang, Z., Qiu, T., Chen, Q., Lu, Y., and Suen, C. Y.
(2021). Document image classification: Progress over
two decades. Neurocomputing, 453:223–240.
Liu, Y., Huang, A., Luo, Y., Huang, H., Liu, Y., Chen,
Y., Feng, L., Chen, T., Yu, H., and Yang, Q. (2020).
Fedvision: An online visual object detection platform
powered by federated learning. In Proceedings of the
AAAI Conference on Artificial Intelligence.
FedBID and FedDocs: A Dataset and System for Federated Document Analysis
557

Love, P. et al. (2013). Document analysis. In Research in
the college context, pages 99–112. Routledge.
Luo, J., Wu, X., Luo, Y., Huang, A., Huang, Y., Liu, Y.,
and Yang, Q. (2019). Real-world image datasets for
federated learning. arXiv preprint arXiv:1910.11089.
Marculescu, M. (2015). Introducing grpc, a
new open source http/2 rpc framework.
https://developers.googleblog.com/2015/02/
introducing-grpc-new-open-source-http2.html.
Mayr, A., Kißkalt, D., Meiners, M., Lutz, B., Sch
¨
afer,
F., Seidel, R., Selmaier, A., Fuchs, J., Metzner,
M., Blank, A., et al. (2019). Machine learning
in production–potentials, challenges and exemplary
applications. Procedia CIRP, 86:49–54.
McMahan, B., Moore, E., Ramage, D., Hampson, S., and
y Arcas, B. A. (2017). Communication-efficient
learning of deep networks from decentralized data. In
Artificial intelligence and statistics, pages 1273–1282.
PMLR.
Paulik, M., Seigel, M., Mason, H., Telaar, D., Kluivers, J.,
van Dalen, R., Lau, C. W., Carlson, L., Granqvist, F.,
Vandevelde, C., et al. (2021). Federated evaluation and
tuning for on-device personalization: System design &
applications. arXiv preprint arXiv:2102.08503.
Polevoy, D. V., Sigareva, I. V., Ershova, D. M., Arlazarov,
V. V., Nikolaev, D. P., Ming, Z., Luqman, M. M.,
and Burie, J.-C. (2022). Document liveness challenge
dataset (dlc-2021). Journal of Imaging, 8(7):181.
Ramaswamy, S., Mathews, R., Rao, K., and Beaufays, F.
(2019). Federated learning for emoji prediction in a
mobile keyboard. arXiv preprint arXiv:1906.04329.
Shoham, N., Avidor, T., Keren, A., Israel, N., Benditkis,
D., Mor-Yosef, L., and Zeitak, I. (2019). Overcoming
forgetting in federated learning on non-iid data. arXiv
preprint arXiv:1910.07796.
S
¨
underhauf, N., Brock, O., Scheirer, W., Hadsell, R., Fox,
D., Leitner, J., Upcroft, B., Abbeel, P., Burgard, W.,
Milford, M., et al. (2018). The limits and potentials of
deep learning for robotics. The International journal
of robotics research, 37(4-5):405–420.
Tan, M. and Le, Q. (2019). Efficientnet: Rethinking
model scaling for convolutional neural networks. In
International conference on machine learning, pages
6105–6114. PMLR.
Xu, M., Liu, J., Liu, Y., Lin, F. X., Liu, Y., and Liu, X. (2019).
A first look at deep learning apps on smartphones. In
The World Wide Web Conference, pages 2125–2136.
Yang, T., Andrew, G., Eichner, H., Sun, H., Li, W., Kong,
N., Ramage, D., and Beaufays, F. (2018). Applied
federated learning: Improving google keyboard query
suggestions. arXiv preprint arXiv:1812.02903.
Zhu, Z., Hong, J., and Zhou, J. (2021). Data-free knowledge
distillation for heterogeneous federated learning. In
International Conference on Machine Learning, pages
12878–12889. PMLR.
VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications
558