Design and Implementation of a Data Model for AI Trustworthiness

Assessment in CCAM

Ruben Naranjo

1,2 a

, Nerea Aranjuelo

1 b

, Marcos Nieto

1 c

, Itziar Urbieta

1,2 d

Javier Fern

andez

3 e

and Itsaso Rodr

ıguez-Moreno

2 f

Vicomtech Foundation, Basque Research and Technology Alliance (BRTA),

Mikeletegi 57, 20009, Donostia-San Sebasti

an, Spain

University of the Basque Country (UPV/EHU), Donostia-San Sebasti

an, Spain

Ikerlan Technological Research Center, Basque Research and Technology Alliance (BRTA),

Jos

e Mar

ıa Arizmendiarreta 2, 20500, Arrasate/Mondrag

on, Spain

{rnaranjo, naranjuelo, mnieto, iurbieta}@vicomtech.org, javier.fernandez@ikerlan.es, itsaso.rodriguez@ehu.eus

Keywords:

AI Trustworthiness, Ontology, Data Model, CCAM, Trustworthiness Assessment.

Abstract:

Amidst the growing landscape of trustworthiness-related initiatives and works both in the academic commu-

nity and from ofﬁcial EU groups, there is a lack of coordination in the nature of the concepts used in these

works and their relationships. This lack of coordination generates confusion and hinders the advances in trust-

worthy AI systems. This confusion is particularly grave in the CCAM domain given nearly all functionalities

related to vehicles are safety-critical applications and need to be perceived as trustworthy in order for them to

become available to the general public. In this paper, we propose the use of a deﬁned set of terms and their

deﬁnitions, carefully selected from the existing reports, regulations, and academic papers; and construct an

ontology-based data model that can assist any user in the comprehension of those terms and their relationship

to one another. In addition, we implement this data model as a tool that guides users on the self-assessment

of the trustworthiness of an AI system. We use a graph database that allows making queries and automating

the assessment of any particular AI system. We demonstrate the latter with a practical use case that makes an

automated trustworthiness assessment based on user-inputted data.

1 INTRODUCTION

In the ever-evolving ﬁeld of transportation, the incor-

poration of Artiﬁcial Intelligence (AI) has emerged as

a crucial factor, especially in the domain of Connected

Cooperative and Automated Mobility (CCAM). As

communities embrace the revolutionary possibilities

of smart transportation systems, the combination of

connectivity, cooperation, and automation has ush-

ered a new era characterised by improved effective-

ness, safety, and convenience (Guerreiro Augusto

et al., 2024). CCAM represents a paradigm shift in

how society perceives and interacts with transporta-

tion. With vehicles seamlessly communicating with

https://orcid.org/0009-0002-3924-0194

https://orcid.org/0000-0002-7853-6708

https://orcid.org/0000-0001-9879-0992

https://orcid.org/0000-0001-7983-3651

https://orcid.org/0000-0002-4867-8115

https://orcid.org/0000-0001-8471-9765

each other and their surroundings, and automated fea-

tures taking centre stage, the new possible mobil-

ity solutions are vast. From improved trafﬁc man-

agement and reduced congestion to improved road

safety and increased accessibility, the impact of AI on

CCAM is far-reaching (Alonso Raposo et al., 2018).

However, amidst this rapid progress, the ethical

dimensions and the need for trustworthy AI in CCAM

cannot be underestimated. As vehicles become more

interconnected and reliant on intelligent decision-

making, ensuring AI systems’ robustness, trans-

parency, and ethical use becomes paramount (Kanak

et al., 2022). Trustworthy AI not only safeguards

the well-being of individuals on the road but also

fosters public conﬁdence in adopting these technolo-

gies. Trustworthiness plays an important role in al-

most any situation facilitated by Information Systems

(IS), especially when uncertainty or potential undesir-

able outcomes are possible (Mcknight et al., 2011).

Various regulations and reports are emerging

around this need (Commission, 2021; Fernan-

Naranjo, R., Aranjuelo, N., Nieto, M., Urbieta, I., Fernández, J. and Rodríguez-Moreno, I.

Design and Implementation of a Data Model for AI Trustworthiness Assessment in CCAM.

DOI: 10.5220/0013153200003890

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

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 3, pages 461-471

ISBN: 978-989-758-737-5; ISSN: 2184-433X

461

dez Llorca and Gomez Gutierrez, 2021; ISO, 2023),

alongside a rapidly growing body of literature (Miller,

2019; Molnar, 2020; Langer et al., 2021; Graziani

et al., 2023). However, these initiatives often lack

coordination, leading to a fragmented approach to

trustworthy AI. The primary hurdle to trustworthi-

ness is establishing a concise deﬁnition of the de-

sired goal. A signiﬁcant challenge lies in the preva-

lence of fuzzy terms and overlapping deﬁnitions sur-

rounding the concept of trustworthiness in different

domains (Graziani et al., 2023). As the integration

of AI in transportation advances, various stakehold-

ers -such as researchers, policymakers, and industry

practitioners- employ different terminologies to de-

scribe and assess the trustworthiness of AI systems

tailored to their needs and perspectives. Along with

this, AI is an area in increasing development; new

subareas, terms, and deﬁnitions arise naturally and

may cause conﬂicts with previous ones. This lack of

standardized language and clear deﬁnitions can lead

to ambiguity, hindering the establishment of univer-

sally accepted benchmarks for trustworthy AI and

slowing down the progress in a common direction.

In this paper, we propose to pave the way towards

a standardised terminology based on existing regu-

lations and literature from various perspectives (e.g.,

technical, legal). Rather than introducing new def-

initions, we focus on connecting existing ones and

deﬁning their relationships in the landscape of trust-

worthy AI. To achieve this, we present an ontology-

based data model that integrates these terms, outlining

their interrelations and requirements. We implement

our data model in a graph database, creating a practi-

cal tool that allows users to evaluate the trustworthi-

ness of AI systems by inputting speciﬁc information

and properties. This tool provides a comprehensive

overview of AI properties in relation to trustworthi-

ness, based on existing initiatives, and regulations.

The main contributions of this paper are:

• We present an ontology that collects and organizes

the most relevant terms related to AI trustworthi-

ness, along with their relationships, in a controlled

vocabulary based on current regulations, reports,

and state-of-the-art of AI trustworthiness.

• We propose a data model to assess the trustwor-

thiness of an AI system for CCAM based on a

speciﬁc set of features provided by the interested

user.

• We offer a tool that guides users through the self-

assessment of the trustworthiness of an AI system.

• We demonstrate the application of our tool with

a practical use case, assessing the trustworthiness

of an AI-based pedestrian detection system.

• We make our data model publicly available to

serve as a reference and support future research

and development in this ﬁeld

2 RELATED WORK

In recent years, there has been an increasingly large

number of publications and regulation initiatives con-

cerning trustworthy AI. The European Government

has been at the forefront of developing a regulatory

framework for AI that is human-centric and trustwor-

thy. The EU AI Act (Commission, 2021) is the ﬁrst

legal framework that sets global standards for AI sys-

tems. It deﬁnes four levels of risk for AI applications:

minimal, limited, high, and unacceptable. Different

regulations and obligations apply to each risk cate-

gory. Minimal risk does not imply any obligation,

limited risk conveys transparency obligations, high

risk involves more speciﬁc requirements, and unac-

ceptable risk leads to prohibited AI systems. The

groundwork for AI regulation was laid with the pub-

lication of the Ethics Guidelines for Trustworthy AI

by the High-Level Expert Group on Artiﬁcial Intel-

ligence (AI HLEG) (Hleg, 2019). This established

legality, ethical adherence, and robustness as the pil-

lars of trustworthiness. These three pillars apply to

the actors and processes involved in the AI systems

(including their development, deployment, and use).

Regarding mobility, the European Joint Re-

search Centre (JRC) published a report on Trust-

worthy Autonomous Vehicles (Fernandez Llorca and

Gomez Gutierrez, 2021) as an application of the ideas

of the AI HLEG to the CCAM sector (high-risk ap-

plications in the EU AI Act). The project reviews

methods and tools to assess the safety, reliability, and

explainability of Autonomous Vehicles (AVs) in real-

world scenarios based on the different SAE levels of

driving automation. It also provides policy recom-

mendations and best practices for deploying AVs in

the EU, taking into account ethical, legal, and social

aspects. This report, along with other European ini-

tiatives (Hennemann et al., 2024), contributes to the

EU’s vision of creating a sustainable, intelligent, and

safe mobility system for all. As can be noticed, the

existing initiatives and regulations expand the dimen-

sions of trustworthiness beyond traditional perspec-

tives that primarily focus on safety to include new as-

pects such as fairness or privacy. Despite these ex-

panded dimensions, safety remains a critical factor.

In the CCAM sector, the use of AI plays a cru-

cial role in performing advanced autonomous vehi-

https://github.com/rnarajo-vicomtech/

ccam-tai-ontology

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

462

cle functionalities, such as visual perception (Perez-

Cerrolaza et al., 2023). These functionalities are of-

ten implemented in safety-related systems, and it is

essential to provide evidence that errors or even the

absence of errors do not lead to system malfunction.

This is achieved by adhering to functional safety stan-

dards, such as those outlined by (IEC, 2010) for in-

dustrial domains and (ISO, 2018), for automotive do-

mains. Consequently, functional safety is one of the

aspects of trustworthiness that must be guaranteed.

However, traditional safety standards were not

originally designed to accommodate AI in safety-

related systems due to their development process,

which relies on probabilistic models generated from

training data. This is unlike traditional software com-

ponents, which are coded from speciﬁcations. Signif-

icant efforts have been dedicated to the adoption of

AI in safety-related systems in recent years (Euro-

pean Union Aviation Safety Agency (EASA), 2023;

Hawkins et al., 2021), application rules (GmbH,

2020), and standards (VDA, 2023; IEC, 2023; IEC,

TBD; ISO/IEC, 2024). Additionally, further stan-

dards are currently under development, such as (ISO,

2023), which will focus on safety and risk factors as-

sociated with AI within the road vehicle context.

However, since these reports were written by poli-

cymakers, law experts, and regulators, there are dis-

crepancies between these reports and academic re-

search on AI systems, such as the use of language and

the end goal of trustworthiness (Gyevnar et al., 2023).

As a matter of fact, there is signiﬁcant debate

and disagreement within the academic community re-

garding the meanings of certain concepts. For in-

stance, some authors use the terms interpretability and

explainability interchangeably (Miller, 2019), while

others give them distinct meanings (Molnar, 2020).

Regarding the interrelations between the terms, we

also ﬁnd discrepancies. For example, some studies

propose AI explainability as a means to support safety

(Jia et al., 2022; Neto et al., 2022), viewing explain-

ability as a requirement for achieving safety. How-

ever, this contrasts with European proposals that treat

explainability and safety as distinct and independent

concepts (Fernandez Llorca and Gomez Gutierrez,

2021; Hleg, 2019). This term deﬁnition and relation-

ship disparity creates confusion and requires careful

attention when reviewing research on these topics.

3 DATA MODEL FOR

TRUSTWORTHY AI IN CCAM

We adopt a three-step methodology to craft our pro-

posed data model for trustworthy AI within the con-

Figure 1: Data Model construction methodology.

text of CCAM. An overview of this methodology can

be seen in Figure 1.

First, we conduct a comprehensive analysis of

current regulations, ofﬁcial reports, and literature re-

lated to trustworthy AI for different stakeholders and

CCAM. This aims to identify miscellaneous yet rele-

vant terms and how these terms interact (Section 3.1).

Leveraging insights from our analysis, we develop

an ontology to collect all relevant terms within their

deﬁnitions and create a data model with the terms’

interrelations and requirements (Section 3.2)

The implementation phase involves translating the

data model into a tangible tool that can be used and

consulted by a user for analyzing the trustworthiness

of an AI system. We use a graph database and a visu-

alization platform for this purpose (Section 3.3).

3.1 Understanding Trustworthiness

The initial step to understanding trustworthy AI is the

deﬁnition of the terminology that forms its founda-

tion. By delineating these terms, we can begin to un-

derstand their relationships that will later be used as a

baseline for constructing our data model.

Regarding our proposed terms, concepts, and def-

initions, we have decided to prioritize ofﬁcial EU

works, as future regulations and laws will be based on

them. Our second priority has been the existing regu-

lations concerning trustworthiness. We also consider

academic research as a third step to address the gaps

left by other works and to obtain an aligned proposal

of all previous domains. The deﬁnitions have been

selected following these criteria: prevalence of terms

in the literature, closeness to the CCAM domain, and

alignment with its meaning outside the AI context.

Design and Implementation of a Data Model for AI Trustworthiness Assessment in CCAM

463

3.1.1 AI System and Properties of an AI System

According to the Council of the Organisation for Eco-

nomic Co-operation and Development (OECD), an

AI system is a machine-based system that can, for

a given set of human-deﬁned objectives, make pre-

dictions, recommendations, or decisions inﬂuencing

real or virtual environments. AI systems are designed

to operate with varying levels of autonomy (OECD,

2019). For example, an increasingly common ADAS

function like Parking Assistance Systems (Naranjo

et al., 2024) is usually considered an AI system be-

cause it needs to make decisions on whether some-

thing around the car is a parking slot, whether it is

occupied, and whether it is suitable for parking, for

which the use of machine learning models is frequent.

Based on the proposals and works analysed we

have identiﬁed the following relevant properties for

AI systems’ trustworthiness:

• Trustworthy: AI system that is beneﬁcial, and is

perceived as beneﬁcial, for the common good of

humanity and the good for the immediate user(s),

improving human welfare and freedom (Fernan-

dez Llorca and Gomez Gutierrez, 2021). In

CCAM environment, it is to note that a trustwor-

thy AI system is considered trustworthy when in-

side its Operational Design Domain which should

be stated to all stakeholders and depends on the

level of vehicle automation (Committee, 2021).

• Lawful: Compliant with all applicable laws and

regulations (Hleg, 2019). This includes laws and

regulations on AI systems as well as laws and reg-

ulations affecting vehicles.

• Ethical: AI system that adheres to ethical princi-

ples and values. It aims to beneﬁt, empower, and

protect both individual human ﬂourishing and the

common good of society (Hleg, 2019).

• Robust: AI system that reliably behaves as in-

tended while minimising unintentional and unex-

pected harm and preventing unacceptable harm to

living beings or the environment. This also ap-

plies to potential changes in its operating environ-

ment or the presence of other agents that may in-

teract with the system in an adversarial manner

(Fernandez Llorca and Gomez Gutierrez, 2021).

• Fair: Assurance that individuals and groups are

free from bias, discrimination and stigmatisation

(Fernandez Llorca and Gomez Gutierrez, 2021).

• Transparent: AI system whose features, compo-

nents, and procedures are open for external in-

spection by a stakeholder. This includes data,

models, algorithms, training methods, decision

mechanisms, and responsibility (ISO/IEC, 2020).

• Safe: AI system that does not lead to a state in

which human life, health, property, or the envi-

ronment is endangered (ISO/IEC, 2022).

• Secure: AI system that has resistance to in-

tentional, unauthorized acts designed to cause

harm or damage to a system (ISO/IEC, 2022).

This includes acts performed physically (Pham

and Xiong, 2021) or through any communication

route such as Vehicle-to-Everything (V2X) com-

munications (Ghosal and Conti, 2020).

• Accurate: AI system whose results of observa-

tions, computations, or estimates are close enough

to the true values or the values accepted as being

true (ISO/IEC, 2022). Accuracy is measured with

different metrics depending on the target AI sys-

tem and AV function. For example, from the per-

spective of overall vehicle performance (e.g., dis-

tance travelled) or from the technology layer per-

spective (e.g., precision and recall in object detec-

tion for scene understanding) (Fernandez Llorca

and Gomez Gutierrez, 2021).

• Reproducible: AI system whose behaviour and

results are consistent when repeated under the

same conditions. (Hleg, 2019).

• Explainable: AI system that produces a clear,

relevant, and accurate presentation of the reason-

ing, functioning, and behaviour behind its output,

using partially or fully automated methods and

directed to clearly deﬁned stakeholders (Gyev-

nar et al., 2023). For instance, a route planning

system should be able to explain the reason be-

hind the route selection. Beyond this, a direct

user of an autonomous driving system, must un-

derstand the state and driving capabilities of the

system (Fernandez Llorca and Gomez Gutierrez,

2021), which might require additional efforts in

terms of Human-Machine Interface (HMI) design.

• Auditable: AI system that enables the assessment

of algorithms, data, and the design process (Fer-

nandez Llorca and Gomez Gutierrez, 2021).

• Accounted For: AI system that has an account-

able stakeholder for each step in its life cycle and

functioning. This Stakeholder is answerable for

actions, decisions and performance of the AI sys-

tem (ISO/IEC, 2022).

• Traceable: AI system whose work items and arte-

facts are uniquely identiﬁable and can be tracked

to the life-cycle step in which they were created

(Fernandez Llorca and Gomez Gutierrez, 2021).

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

464

3.1.2 Stakeholders

Based on the ISO/IEC TS 5723 (ISO/IEC, 2022), a

stakeholder is any individual, group, or organisation

that can affect, be affected by, or perceive itself to

be affected by a decision or activity. The perceived

trustworthiness of an AI system might vary according

to each stakeholder’s needs and interests. We propose

the following uniﬁed subcategories based on (Langer

et al., 2021; Bhatt et al., 2020; Kanak et al., 2022).

• Regulator: Physical or legal entity that designs

and publishes a regulation affecting an AI system.

• Deployer: Physical or legal entity that makes

an AI system available to End-users. Differ-

ent deployer proﬁles include original equipment

manufacturers (OEMs), Tier 1 providers, Tier 2

providers, Tier 0.5 providers, or intermediaries.

• Developer: Physical or legal entity that designs,

builds or maintains an AI system in any of its life-

cycle steps, from sourcing of data to maintaining

or enhancing of the system once it is deployed.

• Domain Expert: Physical or legal entity with ex-

tended knowledge of the domain of a system that

ensures an AI system complies with one or several

regulations, standards, or laws.

• End-User:

– Direct: Physical or legal entity that makes use

of an AI System (e.g. vehicle users).

– Indirect: Physical or legal entity that is af-

fected by an AI System in its functioning cycle.

(e.g. external road users).

3.1.3 Related Terms

In addition to deﬁning the properties of an AI sys-

tem and the stakeholders affected, we identify rela-

tionships between terms and concepts. To allow to

properly build those connections, we need to account

for some additional terms that populate the trustwor-

thiness landscape but are not directly involved with

the AI system or the stakeholders:

• Monitoring Process: Continuous process or

mechanism that tracks provenance, creation, and

evolution of work items and artefacts in the life-

cycle of an AI system, such as data collections, AI

model versions, stakeholders involved, etc.

• KPIs: Key Performance Indicators (KPIs) in the

context of an AI system refer to the quantitative

results and metrics of the AI system.

• Accuracy Requirements: Thresholds that mea-

sure the accuracy of an AI system functions when

compared to KPIs. These are usually established

by regulations, academic research or local laws.

• Audition Process: Process in which a domain ex-

pert reviews and tests the AI system against one or

more regulations and, if successful, emits a certiﬁ-

cate of compliance.

• Regulation: Binding legislative act of general ap-

plication (Council of European Union, 2016).

• Account-Giving Mechanism: Automatic mech-

anism that checks the life-cycle and functioning

status of the AI system and assigns accountability

to the appropriate stakeholder.

• Explanation: Clear, relevant, and accurate pre-

sentation of the reasoning, functioning, and be-

haviour behind the output of an AI system.

• Explainability Mechanism: Mechanism or pro-

cess that generates one or more explanations for

the output of an AI system. There are sev-

eral emerging taxonomies regarding explainabil-

ity; given the lack of standardization, we have

chosen a high-level categorization that organizes

explainability mechanisms depending on strategy.

– Explaining Method: Most commonly known

post-hoc explanations. These methods usu-

ally generate explanations to interpret black-

box models (Hassija et al., 2024). Explana-

tions can be local (for single input values, e.g.

saliency maps) or global (which features con-

tribute more to the overall model).

– Interpretable Design: This is the kind of

mechanism that comes implicitly in inter-

pretable or white-box models. The functioning

of this kind of model is designed to be clear so

that the explanation of the output is the model

itself. e.g. classiﬁcation trees.

– Evaluation Process: Strategy that aims to gen-

erate explanations based on exhaustive testing

of the system in a very extensive set of data

with the objective of predicting the system be-

haviour under different circumstances. This

kind of process is very common in CCAM ap-

plications where auto-makers test functionali-

ties in hundreds or thousands of scenarios.

3.2 Data Model Design

An ontology allows organising the terms related to a

domain, trustworthiness in this case, into a controlled

vocabulary and capturing the relationships between

objects or concepts and the properties that can be used

to describe them. We create an ontology as the base

of our data model.

Design and Implementation of a Data Model for AI Trustworthiness Assessment in CCAM

465

Figure 2: Complete data model. The main classes are in red shades, and the supporting classes are in blue shades.

3.2.1 Elements of the Data Model

We propose AI System and Stakeholder to be the two

main classes acting as pillars for building our data

model. Since the data model’s purpose is to analyze

an AI system’s trustworthiness, the AI system itself

is one of the main building blocks of the data model.

The AI system class has several features or properties

corresponding to some of the terms described in Sec-

tion 3.1 and the risk level deﬁned in the EU AI Act.

These features can be seen in Table 1. It is crucial to

note that although the deﬁned terms (e.g., explainable,

transparent) are general when applied as properties of

the AI System class within the data model, the nota-

tion “<term>To” is employed in certain instances to

specify that the term should consistently relate to a

particular Stakeholder subclass or a group of Stake-

holder subclasses (e.g., ExplainableTo[Developer],

TransparentTo[Direct-user, Deployer]).

The second pillar, Stakeholder, gains importance

when analyzing the relationships between the terms

described in Section 3.1. As some of these terms, e.g.

explainability, are evaluated by different stakeholders,

the Stakeholder itself and its subclasses should be a

key component in analyzing the trustworthiness of the

AI system.

In addition to these pillars, we propose a set

of supporting classes on our data model, that will

help determine the trustworthiness of an AI system.

These supporting classes (e.g., regulation, explain-

ability mechanism) enable the analysis of the AI sys-

tem and its features through the use of some prede-

ﬁned rules. Consequently, these classes are required

to obtain the properties of the AI system. The whole

Table 1: Properties of the AI System and their data types.

Property Data Type Property Data Type

Lawful bool EthicalTo array[string]

Traceable bool Deterministic bool

ExplainableTo array[string] Reproducible bool

Auditable bool Accurate bool

AccountedFor bool Secure bool

TransparentTo array[string] Robust bool

Fair bool TrustworthyTo array[string]

Safe bool RiskLevel integer

set of classes including the two pillars and the sup-

porting classes can be seen in Figure 2.

3.2.2 Rules for Trustworthiness Assessment

Once the data model is deﬁned, we design a set of

rules that assign features to the AI System class based

on user-provided input. These rules are used to eval-

uate the trustworthiness of the AI system deﬁned by

the user.

These rules are the result of our analysis of the re-

lationships of the terms in Section 3.1, and are based

on proposals from the literature, but leave some tasks

that the user shall previously solve. Since this data

model encompasses complex high-level terms such as

fairness, safety, or regulation compliance, the auto-

matic analysis of these elements is out of the scope of

this work. Especially given that the analysis of many

of these concepts is still an open question that is con-

tinually being researched (Parraga et al., 2023; Cav-

iglione et al., 2023; Buyl and De Bie, 2024). The

advances and research in these domains will comple-

ment the work presented in this paper.

The rules are presented summarised in Figure 3

and as pseudo-code in the Appendix. For example,

for a given stakeholder, if the AI system is lawful,

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

466

Lawful

Trustworthy Ethical

Robust

Fair*

Transparent

Safe*

Explainable

AccountedFor

Auditable

Secure*

Accurate

Reproducible Deterministic*

Regulation

Explanation

AuditionProcess

Traceable

Account-giving

mechanism

Monitoring

process

KPIs

Figure 3: Summary of rules for the data model. Each property (in red) entails compliance with other properties or auxiliary

classes of the data model (in blue). Asterisk marks user-inputted data. Note that explainable is for a given stakeholder, and

consequently transparent, ethical, and trustworthy.

ethical, and robust, then the system is trustworthy.

3.3 Data Model Implementation

To bring the data model into practical use, a multi-

step process has been designed to ensure usability and

scalability. This process includes constructing an on-

tology, loading it into a graph database, and enabling

user interaction through a REST API.

• Ontology Construction: the foundation of the

data model implementation begins with construct-

ing the ontology using Prot

, an open-source

ontology editor. The classes, properties, data

types, and relationships in the ontology are de-

ﬁned according to the detailed deﬁnitions pro-

vided in Section 3. This tool allows to metic-

ulously deﬁne and organize the various classes,

properties, and relationships that form the data

model. These elements are deﬁned according to

the detailed deﬁnitions provided in Section 3 of

this paper, which are derived from our extensive

analysis of current regulations and literature.

• Loading the Ontology into the semantic graph

database: once the ontology is constructed, it is

loaded into GraphDB

, a high-performance graph

database optimized for handling ontologies and

linked data. We chose GraphDB for its capa-

bility to store complex relationships and support

SPARQL, a powerful query language used for re-

trieving and manipulating data in RDF format.

• REST API development: To facilitate user in-

teraction with the ontology and database, a REST

API is developed using Flask

, a lightweight web

framework for Python. The REST API acts as an

intermediary, enabling users to seamlessly enter

https://protege.stanford.edu/

https://graphdb.ontotext.com/documentation/10.7/

https://ﬂask.palletsprojects.com/en/3.0.x/

and request data. Internally, the API uses SPARQL

to query the ontology within GraphDB.

4 PRACTICAL APPLICATION OF

THE DATA MODEL TO ASSESS

TRUSTWORTHINESS

This section outlines the tool designed with user-

friendliness in mind, developed as part of the use case

demonstrations in this work. It is decomposed into

two subsections, the ﬁrst one outlines the interface of

the tool and the later the practical application that has

been evaluated in this paper as an example.

Figure 4: Input panel of the assessment tool.

4.1 Tool Interface

Using the database implementation described in Sec-

tion 3.3, we build a graphing and data input tool on

Design and Implementation of a Data Model for AI Trustworthiness Assessment in CCAM

467

Figure 5: Graph visualization and assessment result panels of the assessment tool. Please zoom-in for better visualization.

Grafana

. Grafana is a powerful open-source ana-

lytics and monitoring tool extended with community-

created plugins. We use some of these plugins to add

useful functionality, such as data input with the Data

Manipulation Plugin

and REST API communication

through the Inﬁnity Data Source Plugin

With this setup, we construct a dashboard com-

posed of three main panels:

1. The input data panel allows the user to enter

the required information about their AI system.

This panel uses the Data Manipulation Plugin to

build an input form with the necessary data re-

quired from the user. When the submit button is

clicked, it constructs a POST request and sends it

to the REST API, which communicates with the

database and updates its values.

2. The graph visualization panel displays a graphical

representation of the inputted data and its relation-

ships, following the structure of the data model. It

uses the Inﬁnity Data Source Plugin to perform

GET requests to the REST API, which retrieves

data from the database.

3. The assessment result panel shows the calculated

properties of the AI system using a Table Panel

https://grafana.com/

https://grafana.com/grafana/plugins/

volkovlabs-form-panel/

https://grafana.com/grafana/plugins/

yesoreyeram-inﬁnity-datasource/

from Grafana. The panel uses the Inﬁnity Data

Source Plugin to perform a GET request to the

REST API and displays the properties of the AI

system as a list, showing which of them are ful-

ﬁlled and which of them are not.

4.2 Practical Application

To demonstrate the use of the tool, we show a prac-

tical example in which we want to assess the trust-

worthiness of a pedestrian detection system we have

developed using a Deep Neural Network (DNN).

The ﬁrst step is to enter the required data in the

input data panel (Figure 4). In this panel, the user

shall select and enter data about the AI system to be

assessed. These data are some of the high-level con-

cepts described in Section 3.1, such as whether the

system is fair, deterministic, or has a monitoring pro-

cess. This information should be gathered or analyzed

by the user in a prior step.

In our case, we know that our pedestrian detection

model is deterministic and we have an account-giving

and monitoring process for its entire life-cycle. We

also know that it needs to meet some Average Preci-

sion (AP) requirements, and it does. The model has

an explainability mechanism in place in the form of an

explaining method, in this case, we have a variant of

D-RISE (Petsiuk et al., 2021) that generates explana-

tions of the output that we know our target stakeholder

(direct users) understand. Since we did not manually

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

468

assess some aspects, such as auditability and security,

we select false and/or unknown values in the corre-

sponding columns in the input data panel.

Once the data has been collected, the graph visu-

alization and assessment result panels will be popu-

lated. In them, we can explore the relationships be-

tween the data model classes and the properties of

the trustworthiness landscape that have been calcu-

lated for the assessed AI system. These two panels

are shown in Figure 5.

In the case of our pedestrian detection model, we

can see that it fulﬁls some of the requirements such as

being reproducible and explainable. However, it fails

in other aspects such as transparency and robustness,

requiring the use of additional mechanisms to achieve

a trustworthy system.

5 CONCLUSIONS

In this paper, we presented an ontology that systemat-

ically collects and organizes the most relevant terms

related to AI trustworthiness. Our ontology, based

on current regulations, reports, and state-of-the-art

academic research, provides a controlled vocabulary

that can be universally adopted, enhancing clarity and

consistency in discussions about AI trustworthiness.

Based on the proposed ontology, we built a data

model that can help interested users to evaluate AI

systems based on a well-deﬁned set of features, en-

suring a tailored and relevant assessment process of

the trustworthiness.

We provide a practical tool that facilitates self-

assessment of AI system trustworthiness. The appli-

cability of this tool is demonstrated through a practi-

cal application in a common CCAM domain use case.

This demonstration highlights the tool’s practical util-

ity and effectiveness in real-world scenarios, provid-

ing a concrete example of its application.

In summary, our contributions provide a struc-

tured and practical approach to assessing AI trust-

worthiness, supporting both theoretical understanding

and practical implementation. The work presented

in this paper contributes to the development of open-

source tools, paving the way towards creating trust-

worthy AI systems, with a particular focus on the

CCAM domain.

ACKNOWLEDGEMENTS

This work has received funding from the Basque

Government under project AutoTrust of the program

ELKARTEK-2023, by the European Union’s Hori-

zon Programme under Grant 101076754 — AIthena

Project and by Spanish Ministry of Science and Inno-

vation under project PLEC2023-010240, CAPSUL-

IA.

REFERENCES

Alonso Raposo, M., Grosso, M., Despr

es, J.,

Fern

andez Mac

ıas, E., Galassi, C., Krasenbrink,

A., Krause, J., Levati, L., Mourtzouchou, A., Saveyn,

B., et al. (2018). An analysis of possible socio-

economic effects of a Cooperative, Connected and

Automated Mobility (CCAM) in Europe. European

Union.

Bhatt, U., Xiang, A., Sharma, S., Weller, A., Taly, A., Jia,

Y., Ghosh, J., Puri, R., Moura, J. M., and Eckersley, P.

(2020). Explainable machine learning in deployment.

In Proceedings of the 2020 conference on fairness, ac-

countability, and transparency, pages 648–657.

Buyl, M. and De Bie, T. (2024). Inherent limitations of ai

fairness. Communications of the ACM, 67(2):48–55.

Caviglione, L., Comito, C., Guarascio, M., and Manco, G.

(2023). Emerging challenges and perspectives in deep

learning model security: A brief survey. Systems and

Soft Computing, 5:200050.

Commission, E. (2021). Laying down harmonised rules on

artiﬁcial intelligence (Artiﬁcial Intelligence Act) and

amending certain union legislative acts. Eur Comm,

106:1–108.

Committee, O.-R. A. D. O. (2021). Taxonomy and deﬁni-

tions for terms related to driving automation systems

for on-road motor vehicles. SAE International.

Council of European Union (2016). Treaty on the function-

ing of the european union. article 288.

http://data.europa.eu/eli/treaty/tfeu

2016/art 288/oj.

European Union Aviation Safety Agency (EASA) (2023).

EASA Concept Paper: guidance for Level 1 & 2 ma-

chine learning applications.

Fernandez Llorca, D. and Gomez Gutierrez, E. (2021).

Trustworthy autonomous vehicles. Technical Report

KJ-NA-30942-EN-N (online), Luxembourg (Luxem-

bourg).

Ghosal, A. and Conti, M. (2020). Security issues and

challenges in v2x: A survey. Computer Networks,

169:107093.

GmbH, V. V. (2020). Design and Trustworthiness of au-

tonomous/cognitive systems.

Graziani, M., Dutkiewicz, L., Calvaresi, D., Amorim, J. P.,

Yordanova, K., Vered, M., Nair, R., Abreu, P. H.,

Blanke, T., Pulignano, V., et al. (2023). A global tax-

onomy of interpretable AI: unifying the terminology

for the technical and social sciences. Artiﬁcial intelli-

gence review, 56(4):3473–3504.

Guerreiro Augusto, M., Acar, B., Soto, A. C., Sivrikaya,

F., and Albayrak, S. (2024). Driving into the fu-

ture: a cross-cutting analysis of distributed artiﬁcial

intelligence, CCAM and the platform economy. Au-

tonomous Intelligent Systems, 4(1):1–11.

Design and Implementation of a Data Model for AI Trustworthiness Assessment in CCAM

469

Gyevnar, B., Ferguson, N., and Schafer, B. (2023). Bridging

the transparency gap: What can explainable AI learn

from the AI Act? In 26th European Conference on

Artiﬁcial Intelligence, pages 964–971. IOS Press.

Hassija, V., Chamola, V., Mahapatra, A., Singal, A., Goel,

D., Huang, K., Scardapane, S., Spinelli, I., Mahmud,

M., and Hussain, A. (2024). Interpreting black-box

models: a review on explainable artiﬁcial intelligence.

Cognitive Computation, 16(1):45–74.

Hawkins, R., Paterson, C., Picardi, C., Jia, Y., Calinescu,

R., and Habli, I. (2021). Guidance on the Assurance

of Machine Learning in Autonomous Systems (AM-

LAS).

Hennemann, M., Ebner, G. K., Karsten, B., Lienemann, G.,

and Wienroeder, M. (2024). Data Act: An Introduc-

tion. Nomos Verlagsgesellschaft mbH & Co. KG.

Hleg, A. (2019). Ethics guidelines for trustworthy AI. B-

1049 Brussels.

IEC (2010). IEC 61508(-1/7): Functional safety of elec-

trical / electronic / programmable electronic safety-

related systems.

IEC (2023). ISO/IEC WD 5338 Information technology

— Artiﬁcial intelligence — AI system life cycle pro-

cesses.

IEC (TBD). IEC TS 6254 - Information technology — Ar-

tiﬁcial intelligence — Objectives and approaches for

explainability of ML models and AI systems (Under

Development).

ISO (2018). ISO 26262(-1/11) Road vehicles – Functional

safety.

ISO (2023). ISO/CD PAS 8800 Road Vehicles — Safety

and artiﬁcial intelligence.

ISO/IEC (2020). ISO/IEC TR 24028 - Information tech-

nology — Artiﬁcial intelligence — Overview of trust-

worthiness in artiﬁcial intelligence.

ISO/IEC (2022). ISO/IEC TS 5723 - Trustworthiness —

Vocabulary.

ISO/IEC (2024). ISO/IEC TR 5469 Artiﬁcial intelligence

— Functional safety and AI systems.

Jia, Y., McDermid, J., Lawton, T., and Habli, I. (2022).

The role of explainability in assuring safety of ma-

chine learning in healthcare. IEEE Transactions on

Emerging Topics in Computing, 10(4):1746–1760.

Kanak, A., Erg

un, S., Atalay, A. S., Persi, S., and Karcı,

A. E. H. (2022). A review and strategic approach for

the transition towards third-wave trustworthy and ex-

plainable ai in connected, cooperative and automated

mobility (ccam). In 2022 27th Asia Paciﬁc Conference

on Communications (APCC), pages 108–113. IEEE.

Langer, M., Oster, D., Speith, T., Hermanns, H., K

astner, L.,

Schmidt, E., Sesing, A., and Baum, K. (2021). What

do we want from Explainable Artiﬁcial Intelligence

(XAI)?–A stakeholder perspective on XAI and a con-

ceptual model guiding interdisciplinary XAI research.

Artiﬁcial Intelligence, 296:103473.

Mcknight, D. H., Carter, M., Thatcher, J. B., and Clay, P. F.

(2011). Trust in a speciﬁc technology: An investi-

gation of its components and measures. ACM Trans-

actions on management information systems (TMIS),

2(2):1–25.

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

sights from the social sciences. Artiﬁcial intelligence,

267:1–38.

Molnar, C. (2020). Interpretable machine learning. Lulu.

com.

Naranjo, R., Sintes, J., P

erez-Benito, C., Alonso, P., Del-

gado, G., Aranjuelo, N., and Jevti

c, A. (2024). Park

marking detection and tracking based on a vehicle on-

board system of ﬁsheye cameras. In International

Conference on Robotics, Computer Vision and Intel-

ligent Systems, pages 31–46. Springer.

Neto, A. V. S., Camargo, J. B., Almeida, J. R., and Cug-

nasca, P. S. (2022). Safety assurance of artiﬁcial

intelligence-based systems: A systematic literature re-

view on the state of the art and guidelines for future

work. IEEE Access, 10:130733–130770.

OECD (2019). Recommendation of the Council on Artiﬁ-

cial Intelligence, OECD/LEGAL/044.

Parraga, O., More, M. D., Oliveira, C. M., Gavenski,

N. S., Kupssinsk

u, L. S., Medronha, A., Moura, L. V.,

Sim

oes, G. S., and Barros, R. C. (2023). Fairness in

deep learning: A survey on vision and language re-

search. ACM Computing Surveys.

Perez-Cerrolaza, J. et al. (2023). Artiﬁcial Intelligence for

Safety-Critical Systems in Industrial and Transporta-

tion Domains: A Survey. ACM Comput. Surv.

Petsiuk, V., Jain, R., Manjunatha, V., Morariu, V. I., Mehra,

A., Ordonez, V., and Saenko, K. (2021). Black-box

explanation of object detectors via saliency maps. In

Proceedings of the IEEE/CVF Conference on Com-

puter Vision and Pattern Recognition, pages 11443–

11452.

Pham, M. and Xiong, K. (2021). A survey on secu-

rity attacks and defense techniques for connected

and autonomous vehicles. Computers & Security,

109:102269.

VDA (2023). Automotive SPICE® Process Assessment /

Reference Model Version 4.0.

APPENDIX

Here we present the set of rules designed to make the

self-assessment of an AI system using the proposed

data model. The rules are presented as pseudo-code

in Algorithms 1 to 11.

if AISystem.deterministic then

AISystem.reproducible = True

end

Algorithm 1: Reproducibility rule.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

470

AISystem.accurate = True

foreach KPI in KPIS do

if not (KPI MEETS

AccuracyRequirements) then

AISystem.accurate = False

end

Algorithm 2: Accuracy rule.

if AISystem.isSecure

and AISystem.accurate

and AISystem.reproducible then

AISystem.robust = True

end

Algorithm 3: Robustness rule.

AISystem.law f ul = True

foreach Regulation in AffectingRegulations

if not AISystem COMPLIESWITH

Regulation then

AISystem.law f ul = False

end

Algorithm 4: Lawfulness rule.

if AISystem HAS MonitoringProcess then

AISystem.traceable = True

end

Algorithm 5: Traceability rule.

if AISystem HAS Account-givingMechanism

and AISystem.traceable then

AISystem.accountedFor = True

end

Algorithm 6: Accountability rule.

if AISystem ENABLES AuditionProcess then

AISystem.auditable = True

end

Algorithm 7: Auditability rule.

foreach Stakeholder in Stakeholders do

if AISystem HAS

ExplainabilityMechanism

and Stakeholder UNDERSTANDS

Explanation then

AISystem.explainableTo.append(

Stakeholder)

end

Algorithm 8: Explainability rule.

foreach Stakeholder in Stakeholders do

if AISystem.accountedFor

and AISystem.auditable

and

AISystem.explainableTo[Stakeholder]

then

AISystem.transparentTo.append(

Stakeholder)

end

Algorithm 9: Transparency rule.

foreach Stakeholder in Stakeholders do

if AISystem.fair AND

AISystem.transparentTo[Stakeholder]

AND AISystem.safe then

AISystem.ethicalTo.append(

Stakeholder)

end

Algorithm 10: Ethicalness rule.

foreach Stakeholder in Stakeholders do

if AISystem.lawful

and AISystem.ethicalTo[Stakeholder]

and AISystem.robust then

AISystem.trustworthyTo.append(

Stakeholder)

end

Algorithm 11: Trustworthiness rule.

Design and Implementation of a Data Model for AI Trustworthiness Assessment in CCAM

471