Navigating the Trade-Off Between Explainability and Privacy

Johanna Schmidt

1,∗ a

, Verena Pietsch

2 b

, Martin Nocker

3 c

, Michael Rader

4 d

and Alessio Montuoro

5 e

VRVis Zentrum f

ur Virtual Reality und Visualisierung Forschungs-GmbH, Vienna, Austria

FILL Gesellschaft m.b.H, Gurten, Austria

MCI The Entrepreneurial School, Innsbruck, Austria

Fraunhofer Austria Research GmbH, Wattens, Austria

Software Competence Center Hagenberg GmbH, Hagenberg, Austria

Keywords:

Explainable AI, Privacy, Explainability, Data Visualization, Homomorphic Encryption.

Abstract:

Understanding the rationale behind complex AI decisions becomes increasingly vital as AI evolves. Explain-

able AI technologies are pivotal in demystifying these decisions, offering methods and tools to interpret and

communicate the reasoning behind AI-driven outcomes. However, the rise of Explainable AI is juxtaposed

with the imperative to protect sensitive data, leading to the integration of encryption techniques in AI devel-

opment. This paper explores the intricate coexistence of explainability and encryption in AI, presenting a

dilemma where the quest for transparency clashes with the imperative to secure sensitive information. The

contradiction is particularly evident in methods like homomorphic encryption, which, while ensuring data

security, complicates the provision of clear and interpretable explanations for AI decisions. The discussion

delves into the conﬂicting goals of these approaches, surveying the use of privacy-preserving methods in Ex-

plainable AI and identifying potential directions for future research. Contributions include a comprehensive

survey of privacy considerations in current Explainable AI approaches, an exemplary use case demonstrating

visualization techniques for explainability in secure environments, and identifying avenues for future work.

1 INTRODUCTION

Artiﬁcial Intelligence (AI) has witnessed remarkable

advancements within the last years, revolutionizing

industries and shaping our daily lives. People widely

adopt AI for its unparalleled ability to automate tasks,

derive insights, and offering enhanced efﬁciency and

innovative solutions across diverse industries and ap-

plications. As such, AI algorithms have become a

transformative force across various domains (Saghiri

et al., 2022). For example, AI aids in diagnostics

and personalized healthcare treatments. The ﬁnan-

cial sector beneﬁts from AI for fraud detection and

risk assessment. Mobility applications leverage AI for

trafﬁc management. In manufacturing, AI optimizes

processes through predictive maintenance and quality

https://orcid.org/0000-0002-9638-6344

https://orcid.org/0000-0002-9146-4649

https://orcid.org/0000-0002-6967-8800

https://orcid.org/0009-0000-1819-5246

https://orcid.org/0000-0002-2605-9081

∗

Corresponding author

control. As AI continues to evolve and AI systems

become increasingly sophisticated and integrated into

critical decision-making processes, there is a growing

demand for transparency and accountability, as deci-

sions are made based on intricate patterns and corre-

lations within vast datasets (Zhang and Lu, 2021).

While AI models demonstrate high accuracy in

various tasks, understanding the rationale behind their

choices becomes increasingly crucial. Explainabil-

ity in AI refers to the ability to comprehend and in-

terpret the decisions and actions taken by AI sys-

tems (Barredo Arrieta et al., 2020). Explainable AI

(XAI) technologies are pivotal in demystifying the

decision-making processes of complex algorithms,

addressing the challenges of the “black box” nature

of AI, and mitigating biases. XAI provides methods

and tools to interpret and communicate the rationale

behind AI-driven decisions, making them more un-

derstandable for users and stakeholders. Techniques

range from integrating inherently interpretable mod-

els (Wang and Tuerhong, 2022) to model-agnostic ap-

proaches (Choo and Liu, 2018). Such model-agnostic

726

Schmidt, J., Pietsch, V., Nocker, M., Rader, M. and Montuoro, A.

Navigating the Trade-Off Between Explainability and Privacy.

DOI: 10.5220/0012472200003660

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

In Proceedings of the 19th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2024) - Volume 1: GRAPP, HUCAPP

and IVAPP, pages 726-733

ISBN: 978-989-758-679-8; ISSN: 2184-4321

approaches increasingly use interactive data visual-

ization (Ward et al., 2015) approaches, as these tech-

nologies provide intuitive and transparent means for

understanding complex model decisions, identifying

patterns, and fostering trust among users and stake-

holders (Alicioglu and Sun, 2022).

With increasing interest in explainability, protect-

ing sensitive data became paramount, prompting the

integration of privacy-preserving technologies in AI

development to ensure privacy and security (Timan

et al., 2021). Cryptography protects data from unau-

thorized access, maintaining conﬁdentiality for those

possessing the secret key. While traditional encryp-

tion techniques protect data only during storage or

transmission, Homomorphic Encryption (HE) allows

for computations on encrypted data without prior de-

cryption (Yi et al., 2014). HE has already been suc-

cessfully applied for enabling secure collaboration in

scenarios where data conﬁdentiality is critical, like

Machine-Learning-as-a-Service (Nocker et al., 2023).

HE encrypts model updates that are transferred and

collected at the aggregation server. Federated Learn-

ing (Esterle, 2022) is commonly utilized to collabo-

ratively train an AI model across multiple distributed

data owners while preserving each party’s privacy.

The coexistence of explainability and privacy in

AI poses a complex dilemma (Ogrezeanu et al.,

2022). On the one hand, explainability seeks to de-

mystify the decision-making processes of AI algo-

rithms, enhancing transparency and fostering trust,

involving making the inner workings of models in-

terpretable for users and stakeholders. On the other

hand, privacy-preserving techniques aim to secure

sensitive data and models, rendering them indecipher-

able to unauthorized entities. The contradiction arises

when such methods make providing a clear and inter-

pretable explanation of AI decisions challenging.

In this paper, we discuss the conﬂicting goals of

both approaches. We surveyed the use of privacy-

preserving methods in Explainable AI, an underex-

plored topic, and identiﬁed possible directions for fu-

ture work. Our contributions can be listed as follows:

• A survey on how current XAI approaches have

considered privacy and security.

• Exemplary use case of using visualization tech-

niques for explainability in secure environments.

• Identiﬁcation of directions for future work.

Keeping the balance is crucial for ensuring that AI

systems safeguard sensitive information and maintain

the transparency necessary for user understanding and

acceptance. Addressing this paradox involves explor-

ing innovative hybrid solutions that combine elements

of encryption and explainability.

2 RELATED WORK

Working on Explainable and Privacy-Preserving AI

touches several scientiﬁc directions. A review of dif-

ferent AI technologies and applications (Sheikh et al.,

2023) would be beyond the scope of this paper. In-

stead, we focus on predictive AI, as our use case

in this paper is centered around predictive mainte-

nance, on model-agnostic XAI with a focus on us-

ing visualization techniques, and methods for privacy-

preserving AI.

Predictive AI. AI models for prediction speciﬁcally

aim at making predictions or forecasts based on data

analysis. Predictive AI involves identifying patterns,

trends, and relationships within datasets, allowing

systems to predict future events or outcomes. Sev-

eral types of machine learning algorithms are com-

monly used in predictive AI, each with its strengths

and suitable applications. These algorithms include

approaches from supervised learning like linear and

logistic regression (Lederer, 2021), where temporal

data is used to predict future outcomes. Other su-

pervised learning approaches comprise decision trees

to build tree-like models out of data (Song and Lu,

2015) and support vector machines (SVM) to classify

data points into different categories (Sapankevych and

Sankar, 2009). Random forest, an ensemble learning

approach (Schonlau and Zou, 2020), combines multi-

ple decision trees to make more informed predictions.

Predictive deep learning approaches (Emmert-Streib

et al., 2020) include recurrent neural networks (RNN)

and long short-term memory (LSTM), which handle

sequential data for time series prediction. The choice

of algorithm depends on the data’s speciﬁc character-

istics and the prediction task’s nature. In practice, a

combination of these algorithms, known as ensemble

methods (Maimon and Rokach, 2006), is often em-

ployed to enhance predictive performance. Predictive

AI encompasses a large number of possible applica-

tions. For example, predictive AI plays an important

role in Industry 4.0 for predictive maintenance appli-

cations (Nunes et al., 2023), in ﬁnance for calculating

forecasts (Broby, 2022), and in energy production to

predict energy use (Wang and Srinivasan, 2017). In

this paper we focus on a predictive maintenance

use case.

Explainable AI. Systems and approaches helping

users understand the workings of an AI system can

be summarized as approaches toward XAI. Explain-

ability covers several understandings (Meske et al.,

2022). Explainability can help users understand the

system’s behavior to detect unknown vulnerabilities

Navigating the Trade-Off Between Explainability and Privacy

727

and ﬂaws, bias, and avoid phenomena related to spu-

rious correlations – to evaluate the AI model. Espe-

cially from a developer’s design perspective, under-

standing the inner workings of AI and consequent out-

comes can be vital to increasing the system’s accu-

racy and value with a focus on improvement. From

the end user’s perspective, the application of XAI

can have a positive effect on user trust in the sys-

tem’s decisions (Preece, 2018). Besides inherently

interpretable models (Wang and Tuerhong, 2022),

which we will not cover on this paper, model-agnostic

approaches (Choo and Liu, 2018) provide explana-

tions for the behaviour of existing AI models. Ex-

planations (Yang et al., 2023) can either be source-

oriented (i.e., focusing on input data), representation-

oriented (i.e., connecting input and output), or logic-

oriented (i.e., explaining the inner structure of the

model). All three types of explainability approaches

(source-oriented, representation-oriented, and logic-

oriented) greatly beneﬁt from the use of data visu-

alization to empower humans to explore and under-

stand the results (Samek et al., 2019). Examples for

representation-oriented approaches employing visual-

ization include the Calibrate framework (Xenopoulos

et al., 2023) for the analysis of the output of proba-

bilistic models and approaches for comparing the out-

put of different, similar models (Wang et al., 2023).

As examples for a logic-oriented visualization, the

ActiVis framework (Kahng et al., 2018) aims at vi-

sually displaying neuron activity for model input and

outputs and the LSTMVis (Strobelt et al., 2018) ap-

proach to visualize hidden state dynamics in recurrent

neural networks. Approaches for evaluating and im-

proving model performances also employ visualiza-

tion to make the differences understandable (Jin et al.,

2023). We employed source-oriented visualization

techniques in our use case.

Secure AI. Enhancing data security in AI can be

achieved through robust encryption strategies. File-

level encryption, database encryption, and transpar-

ent data encryption (Gasser and Aad, 2023) work at

the data level and locally secure the data used by the

AI model. End-to-end encryption (Sukhodolskiy and

Zapechnikov, 2020) ensures data protection through-

out its lifecycle. While such encryption techniques

protect data while being transmitted or stored, homo-

morphic encryption (HE) (Yi et al., 2014) is a tech-

nique that enables computations to be performed on

encrypted data without prior decryption.

HE for AI facilitates secure collaborative process-

ing and can be used to secure inference inputs and

the model (Acar et al., 2018). Developing XAI algo-

rithms for models or input data that are homomorphi-

cally encrypted presents signiﬁcant challenges due to

the inherent complexities of this cryptographic tech-

nique. Non-linear operations are not natively sup-

ported by HE, limiting the types of computations that

can be performed in explainability algorithms. Fur-

thermore, HE adds a substantial computational over-

head which makes XAI on encrypted data a resource-

intensive task. As a result, designing and implement-

ing explainability algorithms is a challenging task.

Addressing these challenges is crucial for unlocking

the potential for XAI in scenarios where data conﬁ-

dentiality is paramount.

Federated learning (Zhang et al., 2021) is a ma-

chine learning approach that allows a model to be

trained across decentralized devices or servers hold-

ing local data samples without exchanging them. In

this collaborative learning paradigm, the model is

trained locally on individual devices using local data,

and only model updates (not raw data) are shared with

a central server or aggregator. The central server ag-

gregates these updates to improve the global model,

which is then sent back to the devices. Federated

learning enables the development of machine learning

models without centralizing sensitive data. HE is uti-

lized in federated learning frameworks to hide model

updates from the aggregator or any other party (Phong

et al., 2017). The input data for prediction compu-

tations of the use case were protected using HE.

3 SURVEY: EXPLAINABILITY

VS. SECURITY

We conducted a literature survey to determine

whether existing XAI approaches also address the

dilemma of secure access to data and models and

the need for explainability primarily for end users.

We concentrated on studies, surveys, state-of-the-

art reports, reviews, and books describing general

overviews of XAI approaches in public libraries like

IEEExplore, Google Scholar, Elsevier, and ACM.

Afterward, we studied the publications and checked

whether they addressed the topics security, privacy,

data protection, or encryption. A list and analysis of

the reviewed references can be found in Table 1. All

publications are from recent years, not because we

limited our search but because the XAI topic is still

relatively new in research interest.

Six of the 17 reviewed publications address

the topic of data and model security. Adadi and

Berrada (Adadi and Berrada, 2018) discuss appli-

cations in the ﬁnance industry where data security

and fair lending are to be addressed. Hohman et

al. (Hohman et al., 2019) mention the risk of AI

IVAPP 2024 - 15th International Conference on Information Visualization Theory and Applications

728

Table 1: Reviewed surveys, state-of-the-art reports, and

books on Explainable AI (sorted alphabetically). The ﬁrst

column (Reference) shows the reference to the speciﬁc pa-

per. In the second column (Type) we deﬁne the type of the

publication (S = survey, R = state-of-the-art report or re-

view, B = book). The third column (Sec.) shows whether

the topic of data security and the interplay with explainabil-

ity has been addressed.

Reference Type Sec.

(Adadi and Berrada, 2018) S X

(Alicioglu and Sun, 2022) S –

(Angelov et al., 2021) R –

(Barredo Arrieta et al., 2020) S X

(Brasse et al., 2023) R –

(Chatzimparmpas et al., 2020) R X

(Choo and Liu, 2018) S –

(Da

glarli, 2020) S –

(Danilevsky et al., 2020) S –

(Hohman et al., 2019) S X

(Holzinger et al., 2020) S –

(Liang et al., 2021) S –

(Miller, 2019) S –

(Saeed and Omlin, 2023) S X

(Samek et al., 2019) B X

(Xu et al., 2019) S –

(Yang et al., 2023) S –

models being fooled by inserting wrong data (e.g.,

images) into the training dataset. Saeed and Om-

lin (Saeed and Omlin, 2023) note that model cre-

ators might regard a model’s synthesized knowledge

as conﬁdential and that having only model inputs and

outputs available might be a compromise for still em-

ploying explainability. Together with Barredo Ar-

rieta et al. (Barredo Arrieta et al., 2020), they rec-

ommend further research on XAI tools that explain

model decisions while maintaining a model’s conﬁ-

dentiality (Hohman et al., 2019). In the survey by

Samek et al. (Samek et al., 2019), the authors dis-

cuss the possibility of security breaches when expos-

ing internal details of model structures to outsiders.

They also discuss possible legal implications due to

intellectual property rights since internal information

about models may be proprietary and a fundamental

property right of companies, and the training data may

be privacy sensitive.

A notable void exists concerning the comprehen-

sive integration of robust data security measures in

the realm of XAI. We noticed that there has yet to

be a survey concentrating on the interplay between

explainability, security, and privacy. While XAI solu-

tions aim to enhance transparency and interpretability

in AI decision-making, the focus has often been di-

rected towards unveiling the intricacies of algorithms

rather than fortifying the security of the underlying

data. This oversight raises critical concerns as sensi-

tive information becomes increasingly vulnerable to

potential breaches or unauthorized access. Address-

ing this gap becomes imperative for developing re-

sponsible and trustworthy AI systems.

4 EXPLAINABILITY USE CASE

Embarking on a practical application of predictive AI

and XAI, we describe a practical use case in predic-

tive maintenance and reﬂect on our learnings.

4.1 Use Case Setup

Our investigation focuses on systematically utilizing

HE (Yi et al., 2014) for input privacy in a predictive

analytics use case.

Data and Model. In this use case, a machine learn-

ing model was used to predict the wear of tools in

machine tools to optimize both the service life of tools

and product quality: The inference data and the model

needed to be secured not to reveal intellectual prop-

erty rights, both from the model owner (i.e., the ma-

chine builder) regarding the model and training data

as well as the data owner (i.e., the machine operator)

regarding the inference data and underlying process

know-how. The inference data is encrypted using not

the original data but features computed from the raw

data for training the model (referred to as training fea-

tures). This way, it was also possible to let external

stakeholders use the model by sending only features

(not the raw data) and receiving predictions from the

model. The model remains on the owner’s side and

doesn’t require encryption. Encrypted inputs are re-

ceived by the model owner, who then computes the

output in an oblivious manner.

User Task. The model predicts the wear of tools

based on previous measurements. Process managers

can use the prediction output of the model to deter-

mine the current wear condition of the tool and esti-

mate when the tool should be replaced and whether

the tool life can be extended or should be shortened.

Navigating the Trade-Off Between Explainability and Privacy

729

Explainability. Explainability was especially rele-

vant in this use case because, naturally, high economic

interests are behind it. Replacing parts too early only

partially utilizes the service life of components, some

of which are quite expensive. Replacing components

too late can impair the component’s quality and lead

to unnecessary rejects or damage to other machine

components. End users, therefore, need to know how

much they can trust the prediction to make the right

decisions according to their priorities. We wanted to

give users of the prediction model an impression of

how conﬁdent they can be in the prediction of the

model. Since the model cannot provide any informa-

tion about uncertainties, we decided to focus on the

training features for the analysis. We agreed that users

should be able to compare their input features with the

training features that the model already knows to see

whether the input features roughly correspond to the

situations that the model has already learned. Accord-

ing to Yang et al. (Yang et al., 2023), this describes a

source-oriented approach for explainability.

4.2 Explainability Visualization

Similar to existing approaches (Chatzimparmpas

et al., 2020), we employed data visualization (Ward

et al., 2015) techniques to foster explainability and

communicate model outputs to the end users. We

evaluated three visualization techniques that we in the

following refer to as low, medium, and high explain-

ability. Low explainability refers to providing as little

information about the underlying data and model as

possible, but still providing background information

for users to increase explainability. In the medium

explainability visualization we included more infor-

mation about the data and the model in the visualiza-

tion. The high explainability visualization provides

as much information about the underlying data and

model as possible. A higher level of explainability

comes with a lower security level. We visualized the

input features and training data in an explainability

plot. In all cases, we used a summary bar to the left

of the explainability plot to aggregate the results (i.e.,

how many input data points match the training data).

Low Explainability. To reach low explainability,

we used the training features’ total value range and

displayed this in the background (Figure 1). When

using the low explainability visualization, users could

judge how many input features were outside the train-

ing features (i.e., the known model space). Based on

the colored background, end users could not reveal

the distribution of features in the training dataset.

Medium Explainability. For medium explainabil-

ity, we employed Mean Square Error (Dodge, 2008)

as a similarity measure and mapped the calculated

similarity values to color (Figure 2). We used a cat-

egorical color map with three color bins. This way,

users can compare the similarities of their input fea-

tures with the training data features. End users could

disclose the distribution of the training features from

this visualization, though, when using the visualiza-

tion for sampling with different input features. Af-

ter some iterations, end users could get an impres-

sion of the training feature distribution. Model own-

ers could prevent this by employing technical barri-

ers, e.g., allowing only a certain number of requests

per minute/hour.

High Explainability. The visualization for high ex-

plainability employs density (Figure 3). A density

plot shows all training data points; the input data

points are plotted as dots in the foreground. Users can

use this representation to evaluate the distribution of

the training features and input features together. The

best comparison between input and training features

is possible when end users see the actual distribution

of training features. In this distribution visualization,

outliers and areas of high density are visible. How-

ever, this visualization reveals the values of the un-

derlying training features.

Low Data Variation

High Data Variation

Feature Value

Feature Index

Total

Figure 1: Low Explainability visualization. The training

features’ total value range and displayed this in the back-

ground (light blue). In the foreground, the input features

are shown as bars, either dark blue if inside the training fea-

ture area or red if outside. The bars’ height indicates the

input feature’s value at this position.

IVAPP 2024 - 15th International Conference on Information Visualization Theory and Applications

730

Low Data Variation

High Data Variation

Feature Value

Feature Index

High Similarity

Medium Similarity

Low Similarity

High Similarity

Medium Similarity

Low Similarity

Total

Figure 2: Medium Explainability visualization. We printed

the similarity value of each feature onto the bars. This way

users can analyze the similarity distribution.

4.3 Learnings

In our use case, we could show that discrepancies be-

tween explainability and security still exist. We could

also show that the choice of trust measure impacts

the results (Papenmeier et al., 2022). When analyz-

ing the visualization results in our use case, we could

see that low explainability correlates with reduced se-

curity risks. In contrast, high explainability increases

security vulnerability and privacy leakage, conﬁrming

the known interplay between explainability, security,

and privacy in AI (Ogrezeanu et al., 2022).

For our use case, we clearly identiﬁed medium

explainability as the visualization technique that best

strives the compromise between explainability and se-

curity risk. The similarity measure provides a good

overview of how well the input features match the

knowledge of the model. Securing the model can be

achieved by employing technical barriers.

5 FUTURE WORK

Researchers can contribute to advancing the ﬁeld of

AI, ensuring that future systems are secure, transpar-

ent, interpretable, and aligned with ethical principles,

and continuing future research and development av-

enues. Directions for future work may include:

• Contextual Explainability: Considering con-

textual explainability, where the level of inter-

pretability is adapted based on the context and

Low Data Variation

High Data Variation

Feature Value

Feature Index

Total

Figure 3: High Explainability visualization. A density plot

shows the training data features and the input features are

printed as dots in the foreground.

sensitivity of the data involved, would help to pro-

tect data while still providing a high level of secu-

rity. Such approaches could dynamically adjust

the trade-off between privacy and interpretability,

depending on the speciﬁc requirements of differ-

ent applications.

• User-Controllable Explainability: Investigation

of systems that allow users to control the level of

explainability, e.g., low, medium, or high explain-

ability in our scenario. Users might have varying

preferences for transparency, thus, providing a set

of explainability options can help balance explain-

ability with privacy.

• Education and Awareness: From a user perspec-

tive, fostering education and awareness regarding

the trade-offs between explainability and privacy

would improve the understanding of secure envi-

ronments. Teaching may involve training AI prac-

titioners, policymakers, and the general public to

understand the challenges and potential solutions

in navigating the delicate balance between these

objectives.

• Privacy-Preserving Explainability Techniques:

The development and optimization of explainabil-

ity methods while using privacy-preserving tech-

nologies, e.g., HE, multi-party computation, or

differential privacy. When input data are en-

crypted, the predictive AI computation must be

performed homomorphically encrypted, therefore

also the explainability method, which has not

been addressed enough so far.

Navigating the Trade-Off Between Explainability and Privacy

731

6 CONCLUSION

This position paper addresses the interplay between

explainability, security, and privacy in AI environ-

ments. In conclusion, our use case of an explainability

visualization dashboard for a predictive maintenance

application has illuminated the efﬁcacy of medium

explainability, striking a meaningful compromise be-

tween providing users with interpretable insights and

fortifying the model. These learnings contribute to

the ongoing discourse surrounding the delicate bal-

ance required in navigating the realms of explainabil-

ity and privacy in AI, providing practical implications

for future implementations and emphasizing the need

for tailored approaches in different contexts.

ACKNOWLEDGEMENTS

This work was conducted within the project SMiLe

(No. 4109009) funded by the ’ICT of the Future’ pro-

gram by the Federal Ministry of the Republic of Aus-

tria for Climate Action, Environment, Energy, Mobil-

ity, Innovation and Technology. VRVis is funded by

BMK, BMDW, Styria, SFG, Tyrol, and Vienna Busi-

ness Agency in the scope of COMET (No. 879730),

which is managed by FFG.

REFERENCES

Acar, A., Aksu, H., Uluagac, A. S., and Conti, M. (2018). A

Survey on Homomorphic Encryption Schemes: The-

ory and Implementation. ACM Computing Surveys,

51(4).

Adadi, A. and Berrada, M. (2018). Peeking Inside the

Black-Box: A Survey on Explainable Artiﬁcial Intel-

ligence (XAI). IEEE Access, 6:52138–52160.

Alicioglu, G. and Sun, B. (2022). A survey of visual ana-

lytics for Explainable Artiﬁcial Intelligence methods.

Computers & Graphics, 102:502–520.

Angelov, P. P., Soares, E. A., Jiang, R., Arnold, N. I., and

Atkinson, P. M. (2021). Explainable artiﬁcial intelli-

gence: an analytical review. WIREs Data Mining and

Knowledge Discovery, 11(5):e1424.

Barredo Arrieta, A., D

ıaz-Rodr

ıguez, N., Del Ser, J., Ben-

netot, A., Tabik, S., Barbado, A., Garcia, S., Gil-

Lopez, S., Molina, D., Benjamins, R., Chatila, R., and

Herrera, F. (2020). Explainable Artiﬁcial Intelligence

(XAI): Concepts, taxonomies, opportunities and chal-

lenges toward responsible AI. Information Fusion,

58:82–115.

Brasse, J., Broder, H. R., F

orster, M., Klier, M., and Sigler,

I. (2023). Explainable artiﬁcial intelligence in infor-

mation systems: A review of the status quo and future

research directions. Electronic Markets, 33(1):26.

Broby, D. (2022). The use of predictive analytics in ﬁnance.

The Journal of Finance and Data Science, 8:145–161.

Chatzimparmpas, A., Martins, R. M., Jusuﬁ, I., Kucher, K.,

Rossi, F., and Kerren, A. (2020). The State of the Art

in Enhancing Trust in Machine Learning Models with

the Use of Visualizations. Computer Graphics Forum,

39(3):713–756.

Choo, J. and Liu, S. (2018). Visual Analytics for Explain-

able Deep Learning. IEEE Computer Graphics and

Applications, 38(4):84–92.

Danilevsky, M., Qian, K., Aharonov, R., Katsis, Y., Kawas,

B., and Sen, P. (2020). A Survey of the State of

Explainable AI for Natural Language Processing. In

Proceedings of the 1st Conference of the Asia-Paciﬁc

Chapter of the Association for Computational Lin-

guistics, AACL-IJCNLP ’20, pages 447–459, Suzhou,

China. Association for Computational Linguistics.

glarli, E. (2020). Explainable Artiﬁcial Intelligence

(xAI) Approaches and Deep Meta-Learning Mod-

els. In Aceves-Fernandez, M. A., editor, Advances

and Applications in Deep Learning, chapter 5. Inte-

chOpen.

Dodge, Y. (2008). The Concise Encyclopedia of Statistics.

Springer.

Emmert-Streib, F., Yang, Z., Feng, H., Tripathi, S., and

Dehmer, M. (2020). An Introductory Review of Deep

Learning for Prediction Models With Big Data. Fron-

tiers in Artiﬁcial Intelligence, 3.

Esterle, L. (2022). Deep Learning in Multiagent Systems.

In Iosiﬁdis, A. and Tefas, A., editors, Deep Learning

for Robot Perception and Cognition, pages 435–460.

Academic Press.

Gasser, L. and Aad, I. (2023). Disk, File and Database

Encryption, pages 201–207. Springer Nature Switzer-

land.

Hohman, F., Kahng, M., Pienta, R., and Chau, D. H. (2019).

Visual Analytics in Deep Learning: An Interrogative

Survey for the Next Frontiers. IEEE Trans. on Visual-

ization and Computer Graphics, 25(8):2674–2693.

Holzinger, A., Saranti, A., Molnar, C., Biecek, P., and

Samek, W. (2020). Explainable AI Methods - A Brief

Overview, pages 13–38. ICML ’20. Springer, Vienna,

Austria.

Jin, S., Lee, H., Park, C., Chu, H., Tae, Y., Choo, J., and

Ko, S. (2023). A Visual Analytics System for Im-

proving Attention-based Trafﬁc Forecasting Models.

IEEE Trans. on Visualization and Computer Graph-

ics, 29(1):1102–1112.

Kahng, M., Andrews, P. Y., Kalro, A., and Chau, D. H.

(2018). ActiVis: Visual Exploration of Industry-Scale

Deep Neural Network Models. IEEE Trans. on Visu-

alization and Computer Graphics, 24(1):88–97.

Lederer, J. (2021). Fundamentals of High-Dimensional

Statistics. Springer.

Liang, Y., Li, S., Yan, C., Li, M., and Jiang, C. (2021).

Explaining the black-box model: A survey of local in-

terpretation methods for deep neural networks. Neu-

rocomputing, 419:168–182.

Maimon, O. and Rokach, L. (2006). Data Mining and

Knowledge Discovery Handbook. Springer.

IVAPP 2024 - 15th International Conference on Information Visualization Theory and Applications

732

Meske, C., Bunde, E., Schneider, J., and Gersch, M. (2022).

Explainable Artiﬁcial Intelligence: Objectives, Stake-

holders, and Future Research Opportunities. Informa-

tion Systems Management, 39(1):53–63.

Miller, T. (2019). Explanation in artiﬁcial intelligence: In-

sights from the social sciences. Artiﬁcial Intelligence,

267:1–38.

Nocker, M., Drexel, D., Rader, M., Montuoro, A., and

Sch

ottle, P. (2023). HE-MAN – Homomorphically

Encrypted MAchine Learning with ONnx Models. In

Proceedings of the 8th International Conference on

Machine Learning Technologies, ICMLT ’23, page

35–45, Stockholm, Sweden. ACM.

Nunes, P., Santos, J. P., and Rocha, E. M. (2023). Chal-

lenges in predictive maintenance – A review. CIRP

Journal of Manufacturing Science and Technology,

40:53–67.

Ogrezeanu, I., Vizitiu, A., Cius

del, C., Puiu, A., Coman,

S., Boldis

or, C., Itu, A., Demeter, R., Moldoveanu,

F., Suciu, C., and Itu, L. (2022). Privacy-Preserving

and Explainable AI in Industrial Applications. Ap-

plied Sciences, 12(13):6395.

Papenmeier, A., Kern, D., Englebienne, G., and Seifert, C.

(2022). It’s Complicated: The Relationship between

User Trust, Model Accuracy and Explanations in AI.

ACM Transactions on Computer-Human Interaction,

29(4):35.

Phong, L. T., Aono, Y., Hayashi, T., Wang, L., and Moriai,

S. (2017). Privacy-preserving deep learning via addi-

tively homomorphic encryption. IEEE Transactions

on Information Forensics and Security, 13(5):1333–

1345.

Preece, A. (2018). Asking ‘Why’ in AI: Explainability of

intelligent systems - perspectives and challenges. In-

telligent Systems in Accounting, Finance and Manage-

ment, 25(2):63–72.

Saeed, W. and Omlin, C. (2023). Explainable AI (XAI):

A systematic meta-survey of current challenges and

future opportunities. Knowledge-Based Systems,

263:110273.

Saghiri, A. M., Vahidipour, S. M., Jabbarpour, M. R.,

Sookhak, M., and Forestiero, A. (2022). A Survey

of Artiﬁcial Intelligence Challenges: Analyzing the

Deﬁnitions, Relationships, and Evolutions. Applied

Sciences, 12(8):4054.

Samek, W., Montavon, G., Vedaldi, A., Hansen, L. K., and

uller, K.-R. (2019). Explainable AI: Interpreting,

Explaining and Visualizing Deep Learning. Springer.

Sapankevych, N. I. and Sankar, R. (2009). Time Series Pre-

diction Using Support Vector Machines: A Survey.

IEEE Computational Intelligence Magazine, 4(2):24–

38.

Schonlau, M. and Zou, R. Y. (2020). The Random Forest

Algorithm for Statistical Learning. The Stata Journal,

20(1):3–29.

Sheikh, H., Prins, C., and Schrijvers, E. (2023). Mission AI

- The New System Technology. Springer Cham.

Song, Y.-Y. and Lu, Y. (2015). Decision tree methods: ap-

plications for classiﬁcation and prediction. Shanghai

Archives of Psychiatry, 27(2):130–135.

Strobelt, H., Gehrmann, S., Pﬁster, H., and Rush, A. M.

(2018). LSTMVis: A Tool for Visual Analysis of Hid-

den State Dynamics in Recurrent Neural Networks.

IEEE Trans. on Visualization and Computer Graph-

ics, 24(1):667–676.

Sukhodolskiy, I. and Zapechnikov, S. (2020). Analysis

of Secure Protocols and Authentication Methods for

Messaging. Procedia Computer Science, 169:407–

411.

Timan, T., Mann, Z., Curry, E., Metzger, A., Zillner, S.,

Pazzaglia, J.-C., and Robles, A. G. (2021). Data Pro-

tection in the Era of Artiﬁcial Intelligence: Trends,

Existing Solutions and Recommendations for Privacy-

Preserving Technologies, pages 153–175. Springer Int

Publishing.

Wang, J., Wang, L., Zheng, Y., Yeh, C.-C. M., Jain, S., and

Zhang, W. (2023). Learning-From-Disagreement: A

Model Comparison and Visual Analytics Framework.

IEEE Trans. on Visualization and Computer Graphics,

29(9):3809–3825.

Wang, Y. and Tuerhong, G. (2022). A Survey of Inter-

pretable Machine Learning Methods. In Proceed-

ings of the International Conference on Virtual Re-

ality, Human-Computer Interaction and Artiﬁcial In-

telligence, VRHCIAI ’22, pages 232–237, Changsha,

China.

Wang, Z. and Srinivasan, R. S. (2017). A review of artiﬁ-

cial intelligence based building energy use prediction:

Contrasting the capabilities of single and ensemble

prediction models. Renewable and Sustainable En-

ergy Reviews, 75:796–808.

Ward, M., Grinstein, G. G., and Keim, D. (2015). Interac-

tive data visualization: Foundations, Techniques, and

Applications. AK Peters.

Xenopoulos, P., Rulff, J., Nonato, L. G., Barr, B., and Silva,

C. (2023). Calibrate: Interactive Analysis of Proba-

bilistic Model Output. IEEE Trans. on Visualization

and Computer Graphics, 29(1):853–863.

Xu, F., Uszkoreit, H., Du, Y., Fan, W., Zhao, D., and Zhu,

J. (2019). Explainable AI: A Brief Survey on His-

tory, Research Areas, Approaches and Challenges. In

Proceedings of the CCF International Conference on

Natural Language Processing and Chinese Comput-

ing, NLPCC ’19, pages 563–574, Dunhuang, China.

Springer Int Publishing.

Yang, W., Wei, Y., Wei, H., Chen, Y., Huang, G., Li, X.,

Li, R., Yao, N., Wang, X., Gu, X., Amin, M. B., and

Kang, B. (2023). Survey on Explainable AI: From

Approaches, Limitations and Applications Aspects.

Human-Centric Intelligent Systems, 3:161–188.

Yi, X., Paulet, R., and Bertino, E. (2014). Homomorphic

Encryption, pages 27–46. Springer Int Publishing.

Zhang, C. and Lu, Y. (2021). Study on artiﬁcial intelligence:

The state of the art and future prospects. Journal of

Industrial Information Integration, 23:100224.

Zhang, C., Xie, Y., Bai, H., Yu, B., Li, W., and Gao, Y.

(2021). A Survey on Federated Learning. Knowledge-

Based Systems, 216:106775.

Navigating the Trade-Off Between Explainability and Privacy

733