Human-Centric Dev-X-Ops Process for Trustworthiness in AI-Based

Systems

Antonello Calabr

1 a

, Said Daoudagh

1 b

, Eda Marchetti

1 c

, Oum-El-kheir Aktouf

2 d

and Annabelle Mercier

2 e

Institute of Information Science and Technologies “A. Faedo”, National Research Council of Italy (CNR), Pisa, Italy

Grenoble INP, LCIS - Universit

e Grenoble Alpes, Valence, France

{name.surname}@isti.cnr.it, {name.surname}@grenoble-inp.fr

Keywords:

AI, Agile, Cybersecurity, DevOps, Holistic, Human-Centric, Lifecycle, Privacy, Trustworthiness.

Abstract:

Ai’s potential economic growth necessitates ethical and socially responsible AI systems. Increasing human

awareness and the adoption of human-centric solutions that incorporate, combine, and assure by design the

most critical properties (such as security, safety, trust, transparency, and privacy) during the development will

be a challenge to mitigate and effectively prevent issues in the era of AI. In that view, this paper proposes a

human-centric Dev-X-Ops process (DXO4AI) for trustworthiness in AI-based systems. DXO4AI leverages

existing solutions, focusing on the AI development lifecycle with a by-design solution for multiple desired

properties. It integrates multidisciplinary knowledge and stakeholder focus.

1 INTRODUCTION

Recent researchers estimate the value of AI for the

global economy at around $13 trillion by 2030 (Insti-

tute, 2018) and predict that by 2035, AI could dou-

ble the annual growth rates of gross value added in

12 developed countries (Thaci et al., 2024). With

this enormous potential for economic growth and pos-

sible impact on humans and society, understanding

and promoting AI ethical and social considerations

is urgently needed for every business and manage-

ment (B&M) domain. AI’s ethical and social consid-

erations include fairness, transparency, accountabil-

ity, privacy, bias mitigation, job displacement, and the

broader societal impacts of AI adoption.

As emphasized in the AI Act

, this is especially

critical for AI applications, where data poisoning (i.e.,

the manipulation of data used to train AI models)

and adversarial attacks (i.e., deceiving AI systems by

subtly manipulating inputs that are invisible to hu-

mans but highly inﬂuential to the algorithm) are the

https://orcid.org/0000-0001-5502-303X

https://orcid.org/0000-0002-3073-6217

https://orcid.org/0000-0003-4223-8036

https://orcid.org/0000-0002-0493-9096

https://orcid.org/0000-0002-6729-5590

AI Act can be found at: https://artiﬁcialintelligenceact.

eu/

most common types of attacks. These threats present

a serious risk to the security of AI applications, as

they can compromise the integrity of the results and

lead to harmful consequences, often with signiﬁcant

implications for ethics, privacy, security, and public

trust. As the practice evidences (Song et al., 2017),

to achieve this objective, it is necessary to consider

the above properties to be jointly and by design sat-

isﬁed since the early stages of the development life-

cycle and aligned with social and human needs and

abilities. Therefore, research should move in three

parallel directions:

1. Jointly integrate target properties (like ethics, se-

curity, safety, trust, transparency, and privacy) as

by-design properties of the development lifecycle.

2. Provide new or align existing models, methods,

and tools to the industrial needs and their cost-

saving program.

3. Focus on the needs of the ﬁnal stakeholders (like

ordinary people, companies, organizations, and

governments).

One way to achieve this is by adopting an ethical and

social by-design human-centric development process.

The DevOps development process is one of the most

widely adopted processes that can be easily tailored

to address ethical and social aspects of AI.

A preliminary conceptualization of a DevOps-

based lifecycle for trustable developing systems

288

Calabrò, A., Daoudagh, S., Marchetti, E., Aktouf, O. and Mercier, A.

Human-Centric Dev-X-Ops Process for Trustworthiness in AI-Based Systems.

DOI: 10.5220/0012998700003825

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

In Proceedings of the 20th International Conference on Web Information Systems and Technologies (WEBIST 2024), pages 288-295

ISBN: 978-989-758-718-4; ISSN: 2184-3252

and ecosystems, called 2HCDL (Holistic Human-

Centered Development Lifecycle), has already been

presented in (Daoudagh et al., 2024). Inspired by

this idea and by the shift-left development (Bjerke-

Gulstuen et al., 2015), this paper develops DXO4AI

as a human-centric Dev-X-Ops process for trustwor-

thiness in AI-based systems, where X stands for de-

sired properties (such as ethics, security, safety, trust,

transparency, or privacy).

In particular, the paper provides the following

original contributions: (i) Deﬁnition of the Smart Ob-

jectives (SOs) for targeting the 3 research dimensions

identiﬁed. (ii) Description of the conceived DXO4AI

approach. (iii) Description of an architecture support-

ing DXO4AI. (iv) Preliminary implementation of the

DXO4AI and preliminary results evaluation.

Outline. Section 2 reports on the state-of-the-art.

Section 3 reports the 8 smart objectives we have iden-

tiﬁed. Section 4 describes the conceptualization of

our proposal DXO4AI. Section 5 describes the sup-

porting architecture of the DXO4AI development pro-

cess and its preliminary implementation. Initial re-

sults and discussions are reported in Section 7 and 8.

2 BEYOND THE STATE-OF-THE

ART

AI technology provides development for virtually

endless applications. However, it is often neglected

that ethical vulnerabilities can be exploited through

psychological and social implications via interfer-

ence with human behaviour (Scherr and Brunet,

2017). Even if scholars are becoming aware of the

double-edged sword of technological progress (Win-

ﬁeld et al., 2019), the current practices only aim to

protect humans by considering, for instance, mali-

cious attacks on systems, without considering attacks

that exploit human psychology, overlooking ethics in-

herent to AI systems.

The DXO4AI proposal develops solutions that

leverage societal concerns to the digital evaluation

of technical properties, such as ethics, security, ac-

countability, privacy, etc., that enable self-adaptation

of systems w.r.t. to ethical concerns. Hence, the

DXO4AI goes beyond the current state of the art to re-

spond to the digital disruption caused by societal and

ethical vulnerabilities directed towards undermining

the cohesion and functioning of European industries

and societies. In the DXO4AI proposal, the emer-

gent behavior from AI-based systems is evaluated in

a dedicated Dev-X-Opslifecycle that enables contin-

uous analysis, testing, and evaluation during the de-

sign phase and the gathering of runtime evidence,

w.r.t. properties deﬁning ethical features. This in-

formation can guide the development process toward

continuously improving the system behavior at design

time, preparing it for the runtime operation to support

behavioral adaptation that accounts for the system’s

technical capabilities and social and ethical concerns.

Another aspect close to ethical and social con-

cerns when using digital technologies, including AI-

based systems, is sustainability. This is addressed by

the European “Green Deal” (to the European Parlia-

ment, 2019) and will be considered in the DXO4AI

proposal through the targeted use case, which is re-

lated to developing an intelligent decentralized sys-

tem to model collaboration between human operators

and drones in wildﬁre ﬁghting. The chosen underly-

ing intelligent model is a multi-agent system. Indeed,

the multi-agent paradigm is particularly suited to de-

ploying intelligent and autonomous systems (Cal-

varesi et al., 2017; Dorri et al., 2018). Such systems

are found in many new applications based on intelli-

gent nodes placed in natural environments or close to

users to measure, optimize, and reduce resource us-

age. For example, optimizing the energy consump-

tion in a building (Hafsi et al., 2021), or responding

to climate issues (Bibri et al., 2024).

Also, DXO4AI will enhance past projects’ inno-

vative methods and tools with human-centricity. This

assures trustworthiness in multiple directions, includ-

ing technical and social directions, boosting the gen-

eral level of trust in emergent new digital develop-

ments and boosting the technological adoption of in-

novative solutions.

3 ENVISIONED SMART

OBJECTIVES

By focusing on the realization of the three parallel

research directions identiﬁed in the previous section,

the following Smart Objectives (SOs) should be con-

sidered.

Holistic Approach (SO1). The complexity of AI-

based systems and applications and the diversity of

the stakeholders involved in the conceiving, develop-

ment, implementation, and use require holistic solu-

tions to consider all the system dimensions: software,

hardware, automation, electronics, and corresponding

stakeholders’ expertise and knowledge, in addition to

social and ethical requirements (Thomas et al., 2019).

Human-Centered Approach (SO2). Supporting

human-centered development in AI is essential for

aligning with social and ethical values, sustainabil-

Human-Centric Dev-X-Ops Process for Trustworthiness in AI-Based Systems

289

ity, and trustworthiness. Enhancing multidisciplinary

stakeholder involvement throughout the AI develop-

ment lifecycle promotes public awareness, adoption

of AI methods, and transparency. The Internet of Peo-

ple (IoP) (Miranda et al., 2015) is a recent data man-

agement paradigm that helps model and predict mis-

behavior or accidents.

Modeling the Behavior (SO3). Behavioral proﬁles

of stakeholders in a target application domain should

be considered throughout systems’ modeling, imple-

mentation, validation, and prediction (Dobaj et al.,

2022). AI, Digital Twins, crowdsourcing, and col-

laborative platforms can help create these proﬁles.

These proﬁles should incorporate understanding re-

lationships by combining various functional and non-

functional aspects.

Integrated By-Design Approach (SO4). Promot-

ing ”by-design” approaches, such as Privacy By De-

sign (Cavoukian et al., 2009)), is increasingly becom-

ing a legal obligation, exempliﬁed by the GDPR’s

Data Protection By Design and By Default (Art.

25) (Commission, 2016). These principles should be

integrated early in development to prevent ﬂaws, vul-

nerabilities, and issues due to new devices and com-

ponents.

Self-Adaptation and Prediction (SO5). Self-

adaptive methodologies that ensure components and

devices are trustable, validated, and veriﬁed before in-

tegration into complex environments help reduce de-

velopment costs in case of problems (Casimiro et al.,

2021). Frameworks for measurable, risk-based trust

to develop, deploy, and operate complex, intercon-

nected ICT systems are important for smart failure

predictions (Calabr

o et al., 2024).

Multidisciplinary Approach (SO6). Different

sources of knowledge and requirements, such as the

Law (e.g., regulation and directives), standards, tech-

nical speciﬁcations, and domain-speciﬁc best prac-

tices, should be taken into account from the beginning

to derive a set of technical requirements that can be

used for developing the intended digital solutions that

will consider social and ethical properties (Thomas

et al., 2019).

Quantitative and Qualitative Proposal and Solu-

tions (SO7). Effective and efﬁcient development ne-

cessitates quantitative and qualitative data collection

and analysis methods]. Proposals should incorpo-

rate risk management and prevention; capabilities for

modeling, testing, monitoring, and analyzing cyberse-

curity risks, attacks, and violations; and stakeholder-

driven, domain-speciﬁc requirements (Van Looy,

2021). They should adopt standards, metrics, guide-

lines, and approaches to ensure functional proper-

ties such as security, safety, trust, transparency, and

privacy throughout the entire lifecycle. Integrating

these standards and metrics is essential for maintain-

ing these properties over the system’s lifetime.

Combining Different Xs (SO8). Non-functional

requirements, known as Xs properties (such as ac-

countability, trust, privacy, security, safety, and

transparency), have traditionally been studied sepa-

rately (Giraldo et al., 2017). However, in many con-

texts, these properties can interact deeply or conﬂict.

Integrated approaches are needed to combine and an-

alyze the Xs properties during system execution to

achieve the required quality. According to the liter-

ature, the relationships between these properties can

be classiﬁed into (Kriaa et al., 2015): (1) Indepen-

dence, where Xs are deﬁned independently; (2) En-

forcement, where Xs impact each other through con-

ditional relationships, mutual reinforcement, or mu-

tual overlap; and (3) Conﬂicts, where antagonisms or

oppositions exist between two or more Xs.

4 IMPLEMENTATION

GUIDELINES

The implementation of the SOs mentioned above will

be performed considering the following guidelines:

Multidisciplinary. This dimension is found in two

main aspects: (1) exploiting and integrating differ-

ent sources of information and knowledge and (2)

promoting collaboration with experts in the require-

ments, analysis, design, deployment, and runtime

phases (as suggested by SO2 and SO3).

To Be Holistic. The DXO4AI proposal applies to the

development of the system, the ecosystem, and their

hardware-software (HW-SW) components, consider-

ing all system dimensions (software, hardware, au-

tomation, electronics) (as suggested by SO1). It also

targets vulnerabilities, erroneous state detection, and

satisfaction of different application domains’ needs,

requirements, and properties (see SO7).

To Be Human-Centric. As suggested by SO2, it pro-

vides easy-to-use facilities focusing on the commonly

adopted technologies that put humans, social interac-

tion, and ethical concerns at the core of digital and

AI services. It adopts the “Internet of People (IoP)”

paradigm. Additionally, as suggested by SO7, stake-

holders will continuously have a clear view of what

is going on, be able to verify properties or express

needs, get certiﬁcation and assurance of the process

and the applied methodology, get the continuous en-

forcement of the required properties; and, be able to

exercise their rights.

To Be Focused on X-Aware Properties (-X-).

As suggested by SO8, it provides the possibil-

WEBIST 2024 - 20th International Conference on Web Information Systems and Technologies

290

Figure 1: Human-centric Dev-X-Ops Process for Trustworthiness in AI-based Systems (DXO4AI).

ity to combine and assures Security (X=Sec), Pri-

vacy (X=Pri), Transparency (X=Tra), Lawfulness

(X=Law), Accountability (X=Acc), as well as Au-

ditability (X=Aud) and Certiﬁcation (X=Cer). Addi-

tionally, as suggested by SO2 and SO7, it provides

a means for sharing responsibilities throughout the

entire system and ecosystem lifecycle, assuring the -

X- properties for leveraging consciousness, learning,

shared knowledge, and overall Quality.

To Support Continuous and Incremental Delivery.

As suggested by SO4, the DXO4AI proposal interre-

lates two main phases, development, and operation,

for continuous delivery and mutual feedback. As in-

dicated by SO7, the DXO4AI also promotes the in-

cremental adoption of and compliance with standards,

metrics, and guidelines throughout the lifetime.

To Be Based on By-Design Principles. As suggested

by SO4, it includes the X-by-design principles (such

as Security-by-Design and Privacy-by-Design) in all

the development and operational stages. The possi-

bility of customizing the life cycle depending on the

combination of Xs (as suggested by So8), the appli-

cation domain environment, and the stakeholders’ be-

havioral proﬁles (as indicated by SO3) are pivotal el-

ements of the novelty of the DXO4AI proposal.

To Support Self-Adaptation and Timely Predic-

tion. According to SO5, monitoring and logging en-

hanced with X-based technologies can be essential in

assuring self-management and assessment of systems

and ecosystems. Continuous and incremental delivery

(as suggested by SO4) and using behavioral proﬁles

(as indicated by S03) can provide a clear understand-

ing of Xs violations and threats.

5 DEVELOPMENT PROCESS

By leveraging the preliminary proposal of 2HCDL

(Daoudagh et al., 2024), the DXO4AI concep-

tual process, depicted in Figure 1, includes two

phases: the Holistic Human-Centric Development

phase (DXO4AI Dev) and the Holistic Human-Centric

Operation phase (DXO4AI Ops). Therefore, the pro-

cess is transformed into “Dev-X-Ops methodology,”

where the X represents the X (or combination of) de-

sired nonfunctional property for each target system,

ecosystem, or constituent HW/SW component. The

realization of the DXO4AI Dev phase will be guided

by analyzing sources of knowledge (e.g., speciﬁca-

tion of vulnerabilities, law and EU directives, system,

and ecosystem speciﬁcation) and include three steps:

Modeling, X-by-Design Development, and Valida-

tion. In particular, different proposals and methodolo-

gies will be considered when realizing each step. For

instance, the Modelling step could use Model-Driven

Architecture (MDA) or Semantic Web-based solu-

tions (such as Ontologies) (Daoudagh et al., 2023).

During the realization of the DXO4AI Ops phase, dif-

ferent sources of information will be considered to

deﬁne its three steps: Deployment, Monitoring and

Logging, and Reports & Recommendations.

6 PRELIMINARY

IMPLEMENTATION

Figure 2 and Figure 3 present the supporting architec-

ture of the DXO4AI development process explained

in the previous section. They realize the DXO4AI

Dev and DXO4AI Ops phases described in Figure 1,

respectively. In ﬁgures, the human in the center

can play different roles, e.g., tester, developer, le-

gal expert, user, cybersecurity expert, or data protec-

tion ofﬁcer. The preliminary implementation of the

DXO4AI architecture relies on several existing arti-

facts that collaborate through a supporting framework

that accommodates the DXO4AI Dev and DXO4AI

Ops phases (Daoudagh et al., 2024).

6.1 DXO4AI Dev Implementation

The main components described for the DXO4AI Dev

are realized by leveraging and composing the follow-

ing existing artifacts.

Human-Centric Dev-X-Ops Process for Trustworthiness in AI-Based Systems

291

Figure 2: DXO4AI: Proposed Architecture Supporting Dev

Phase.

Knowledge Management encompasses the man-

agement of different sources of information, e.g., in-

dustrial standards, speciﬁc databases like the CVE

and CWE databases for security analysis, etc. Lever-

aging machine-readable representations, such as on-

tologies, DXO4AI allows an automated adaptation of

mitigation solutions related to a particular X or a com-

bination of X properties in the presence of identiﬁed

threats. The implementation of this component relies

on the domain-based ontology DAEMON (Daoudagh

et al., 2023) and supports relationships among SoS,

IoT (Calabr

o A., 2021).

User/Domain Customization of X-USs focuses

on managing user interaction and providing action-

able User Stories (X-USs). X-USs are machine-

readable representations of the desired non-functional

properties (Xs) that users can select and customize.

This allows developers to incorporate user needs and

domain-speciﬁc considerations into the development

process—the implementation. DXO4AI relies on

GDPR-based User Stories deﬁned in (Bartolini et al.,

2019) and organized in speciﬁc Data Protection back-

logs, which are lists of User Stories about GDPR pro-

visions told as technical requirements.

Modeling & Coding component integrates vari-

ous modeling approaches, such as UML diagrams and

Domain-Speciﬁc-Languages (DSLs), to provide valu-

able support for the coding phase. By leveraging be-

havioral models, the dev and ops phases will mutu-

ally enrich each other. In implementing this compo-

nent, two open-source tools for behavioral modeling

are under evaluation: Xtext

and ANTLR

. Xtext

empowers developers to design DSLs speciﬁcally tai-

lored to a speciﬁc domain. These DSLs enable the

creation of concise and readable models that effec-

tively capture the system’s behavior. This focus on

clarity within a particular domain makes Xtext valu-

https://projects.eclipse.org/projects/modeling.tmf.xtext

https://www.antlr.org/

able for DXO4AI. ANTLR (ANother Tool for Lan-

guage Recognition) has a different but complemen-

tary strength. While not directly generating code,

ANTLR allows building parsers and interpreters for

the custom DSLs created with Xtext. This unique

combination unlocks the potential to create highly

readable, executable models.

Testing & Validation focuses on speciﬁc Xs un-

der evaluation and allows the integration of different

tools and approaches for continuously assessing the

system’s properties during development. In the cur-

rent implementation, a broader security and privacy

testing Toolbox is speciﬁcally designed for access

control systems, considering GDPR compliance that

includes: (1) XACMET (XACML Modeling & Test-

ing) tackles two essential tasks: generating XACML

requests (used in access control) and acting as an au-

tomated oracle to measure test coverage (Daoudagh

et al., 2020); (2) XACMUT (XACml MUTation) fo-

cuses on generating variations (mutants) of XACML

policies (Bertolino et al., 2013); and (3) GROOT pro-

vides a unique methodology for combinatorial testing

and helps evaluate compliance with the GDPR and its

contextualization within a target system (Daoudagh

and Marchetti, 2021).

6.2 DXO4AI Ops Implementation

The main components of the DXO4AI Ops phase are

described in the following and are realized by lever-

aging and composing existing artifacts as follows.

Figure 3: DXO4AI: Proposed Architecture Supporting Ops

Phase.

Operational Environment Setting aims at se-

lecting X-based behavioral models useful for self-

assessment or predicting possible violations and

threats during the operation phase. This is also in

charge of deﬁning the operational test data if neces-

sary. The component relies on FIISS (Priyadarshini

et al., 2023), which analyses a target system’s ar-

chitectural and behavioral speciﬁcations to identify

safety and security interactions. These implementa-

WEBIST 2024 - 20th International Conference on Web Information Systems and Technologies

292

tions contribute to realizing the methodology by inte-

grating and extending them to cover other Xs proper-

ties. The DXO4AI project will allow more in-depth

research on social and ethical issues regarding this

methodology and the underlying preliminary.

Operational Proﬁle Deﬁnition aims at selecting

X-based behavioral models useful for self-assessment

or predicting possible violations and threats during

the operation phase. This is also in charge of deﬁn-

ing the operational test data if necessary.

Operational Environment Setting sets up the

operational environment and speciﬁes the required in-

strumentation for monitoring and reporting activities.

Monitoring & Logging collects data during the

operation, assesses the Xs properties, and launches

necessary countermeasures in case of detected viola-

tions or misbehavior. The component is implemented

through the Concern Monitoring Infrastructure

. It

is an open-source, customizable, and generic moni-

toring proposal that has already been evaluated as ap-

propriate in several speciﬁc contexts and application

domains (such as (Calabr

o and Marchetti, 2024; Cal-

abr

o et al., 2016)) for evaluating functional and non-

functional properties. Additionally, to allow loosely

coupled communication and to manage vast amounts

of data, the Concern communication backbone imple-

mentation is message-based and ready for integration

through REST interfaces.

Data analytics performs post-analysis of the data

collected during the operation execution, suggests

countermeasures in case of Xs violations, and im-

proves for successive development iterations.

7 PRELIMINARY RESULTS

The multi-agent system paradigm has gained interest

with the widespread adoption of IoT and AI-based

systems and their need for intelligence (reactiveness

and proactiveness). The embedded multi-agent sys-

tem model combines hardware and software com-

ponents, supporting various applications such as au-

tonomous vehicles and smart grids. However, this

model increases the need for trust among agents, as

some could be malicious and intend to harm the entire

system’s operation. Trust is a property that can pri-

marily beneﬁt from the DXO4AI methodology. Con-

sidering the peculiarity of the trust management sys-

tems, DXO4AI methodology focuses on the follow-

ing actions: (1) Information gathering for trust evalu-

ation using evidence from past interactions, contexts,

and other agents; (2) Trust modeling and evaluation

https://github.com/ISTI-LABSEDC/Concern

to represent trust in an agent. It describes how trust-

related values are deﬁned and calculated from evi-

dence; and (3) Decision-making to evaluate the effect

of decisions made on trust values.

DXO4AI methodology has been applied to a trust

management proof-of-concept system (PoC) repre-

senting traditional applications for explorers and har-

vester agents (Darroux et al., 2019). Speciﬁcally, a

group of light and rapid agents are explorers and in-

vestigate targeted resources disseminated in a given

ﬁeld. Once explorer agents ﬁnd resources, harvester

agents are informed of the resources’ locations to

bring them back to a base. An issue may arise from

malicious explorer agents, which can indicate incor-

rect locations, causing harvester agents to run out

of energy. By applying the DXO4AI methodology

and instantiating it to trust (X=trust), the evolution of

trust among a group of explorer and harvester agents

is evaluated by considering some malicious explorer

agents. This PoC focuses on three different potential

behaviors of harvester agents concerning their inter-

actions with explorer agents:

• naive behavior: the agent only uses experience

to adapt its trust level. This behavior is more

straightforward when most explorers are trustwor-

thy because the likelihood of getting reliable in-

formation is high. But if the harvester agent se-

lects a malicious explorer agent, it loses part of its

energy needed to go to the resource location and

return without any resources.

• cooperative behavior: in his behavior, the har-

vester agent will select the explorer agent using

his own experience and ask for recommendations

from other agents.

• basic learning behavior: in this behavior, the har-

vester agent can choose between the naive behav-

ior and the cooperative behavior based on avail-

able trust information. The learning process is

based on the multi-arm bandit model (Xia et al.,

2017)

For harvester agents, simulations are deployed us-

ing three agents’ behaviors (naive, cooperative, and

MABTrust). They operate in four different environ-

ments where the number of malicious agents differs

to see how trust inﬂuences the overall result. The sim-

ulations were done with 80 agents with:

1. no malicious agents: 50 reliable explorer agents

and 30 harvesters (50e);

2. 40% of malicious explorer agents: 30 reliable ex-

plorer agents, 20 malicious explorer agents, and

30 harvesters (30e20m);

3. 50% of malicious explorer agents: 25 reliable ex-

plorer agents, 25 malicious explorer agents, and

Human-Centric Dev-X-Ops Process for Trustworthiness in AI-Based Systems

293

Figure 4: Results with different conﬁgurations: (a) no malicious (b) 40% malicious(c) 50% malicious (d) 60% malicious.

30 harvesters (25e25m);

4. Most malicious explorer agents (60%): 20 re-

liable explorers, 30 malicious explorers, and 30

harvesters (20e30m).

Figure 4 shows the resources the harvester agents col-

lected before running out of energy. In an environ-

ment without malicious explorer agents (a), the MAB

algorithm does not perform as well as the naive or co-

operative behavior. This is because, in this environ-

ment, agents ought to use most of their energy to col-

lect resources and not as much for the Trust Manage-

ment System (TMS) because it will reduce their per-

formance. Within environments 30e20m and 25e25m,

the MAB algorithm performs better than the others.

In the last conﬁguration (d), with the most trustworthy

harvester agents, the cooperative behavior performs

better than the two others, with the MAB being a close

second. Using the DXO4AI methodology, we ana-

lyzed the trust property within an intelligent system

under development. This process allows for the trust

model to be adjusted for future design cycles.

8 EXPECTED OUTCOMES AND

DISCUSSION

Even if in the proposal stage, the presented smart ob-

jectives, the DXO4AI methodology, and its prelim-

inary supporting architecture envision different im-

pacts and outcomes for the research and industrial en-

vironment. Indeed, they can stimulate the research

in the design and development of speciﬁc models and

methods and underlying platforms and tools for en-

forcing the by-design and combined implementation

of Xs properties during the development and opera-

tion phases. Leveraging the DevOps principles, the

proposed solution will be an industrial, practical, and

effective approach for continuously assessing and en-

hancing the considered properties throughout the en-

tire lifecycle. Finally, the developed processes, mod-

els, and tools could lead to the proposal of patents

or licensed platforms, increasing economic/industrial

impact. Considering the human aspects, the DXO4AI

could contribute to leveraging the Xs awareness and

education in general for any possible stakeholder

(both professionals and ordinary users) with a sub-

stantial societal impact. DXO4AI can close the cur-

rent literature and technological gap in combining se-

curity and privacy by design (Abu-Nimeh and Mead,

2012). We expect our approach to be generic enough

to consider other processes and properties. The novel

Integrated framework will allow the industrial con-

text to integrate processes such as threat analysis, risk

analysis, testing, formal veriﬁcation, effective audit

procedures for cybersecurity testing, validation, and

consideration of certiﬁcation aspects. It also promotes

the behavioral model as an effective means of model-

ing and testing Xs properties and analyzing HW/SW

components to discover their potential vulnerabilities.

ACKNOWLEDGEMENTS

This work was partially supported by the project

RESTART (PE00000001), the project SER-

ICS (PE00000014), and the project THE (CUP

B83C22003930001) under the NRRP MUR program

funded by the EU - NextGenerationEU.

REFERENCES

Abu-Nimeh, S. and Mead, N. R. (2012). Combining se-

curity and privacy in requirements engineering. In

Information Assurance and Security Technologies for

Risk Assessment and Threat Management: Advances,

pages 273–290. IGI Global.

Bartolini, C., Daoudagh, S., Lenzini, G., and Marchetti, E.

(2019). Gdpr-based user stories in the access control

perspective. In QUATIC 2019, Ciudad Real, Spain,

September 11-13, 2019, Proceedings, pages 3–17.

Bertolino, A., Daoudagh, S., Lonetti, F., and Marchetti, E.

(2013). XACMUT: XACML 2.0 mutants generator.

In ICST 2013 Workshops Proceedings, Luxembourg,

Luxembourg, March 18-22, 2013, pages 28–33. IEEE

Computer Society.

Bibri, S. E., Krogstie, J., Kaboli, A., and Alahi, A. (2024).

Smarter eco-cities and their leading-edge artiﬁcial in-

telligence of things solutions for environmental sus-

tainability: A comprehensive systematic review. En-

vironmental Science and Ecotechnology, 19:100330.

WEBIST 2024 - 20th International Conference on Web Information Systems and Technologies

294

Bjerke-Gulstuen, K., Larsen, E. W., St

alhane, T., and

Dingsøyr, T. (2015). High level test driven

development–shift left. In XP 2005 Conference, pages

239–247. Springer.

Calabr

o, A. and Marchetti, E. (2024). MOTEF: A test-

ing framework for runtime monitoring infrastructures.

IEEE Access, 12:38005–38016.

Calabr

o A., Daoudagh S., M. E. (2021). Mentors: Moni-

toring environment for system of systems. In WEBIST

2021, pp. 291–298, 26-28/10/2021.

Calabr

o, A., Daoudagh, S., and Marchetti, E. (2024). To-

wards enhanced monitoring framework with smart

predictions. Log. J. IGPL, 32(2):321–333.

Calabr

o, A., Lonetti, F., Marchetti, E., and Spagnolo,

G. O. (2016). Enhancing business process perfor-

mance analysis through coverage-based monitoring.

In QUATIC 2016, Lisbon, Portugal, September 6-9,

2016, pages 35–43. IEEE Computer Society.

Calvaresi, D., Marinoni, M., Sturm, A., Schumacher, M.,

and Buttazzo, G. (2017). The challenge of real-time

multi-agent systems for enabling iot and cps. In Pro-

ceedings of the international conference on web intel-

ligence, pages 356–364.

Casimiro, M., Romano, P., Garlan, D., Moreno, G. A.,

Kang, E., and Klein, M. (2021). Self-adaptation for

machine learning based systems. In ECSA (Compan-

ion).

Cavoukian, A. et al. (2009). Privacy by design: The 7 foun-

dational principles. Information and privacy commis-

sioner of Ontario, Canada, 5:2009.

Commission, E. (2016). Regulation (EU) 2016/679 of the

European Parliament and of the Council of 27 April

2016 (General Data Protection Regulation). Ofﬁcial

Journal of the European Union, L119:1–88.

Daoudagh, S., Lonetti, F., and Marchetti, E. (2020).

XACMET: XACML testing & modeling. Softw. Qual.

J., 28(1):249–282.

Daoudagh, S. and Marchetti, E. (2021). Groot: A gdpr-

based combinatorial testing approach. In ICTSS 2021,

London, UK, November 10–12, 2021, Proceedings,

page 210–217, Berlin, Heidelberg. Springer-Verlag.

Daoudagh, S., Marchetti, E., and Aktouf, O.-E.-K. (2024).

2hcdl: Holistic human-centered development lifecy-

cle.

Daoudagh, S., Marchetti, E., Calabr

o, A., Ferrada, F.,

Oliveira, A., Barata, J., Peres, R. S., and Marques, F.

(2023). DAEMON: A domain-based monitoring on-

tology for iot systems. SN Comput. Sci., 4(5):632.

Darroux, A., Jamont, J.-P., Aktouf, O.-E.-K., and Mercier,

A. (2019). An energy aware approach to trust manage-

ment systems for embedded multi-agent systems. In

Software Engineering for Resilient Systems - Serene

2019 workshop, pages 121–137.

Dobaj, J., Riel, A., Krug, T., Seidl, M., Macher, G.,

and Egretzberger, M. (2022). Towards digital twin-

enabled devops for cps providing architecture-based

service adaptation & veriﬁcation at runtime. In

SEAMS, SEAMS ’22, page 132–143, New York, NY,

USA. Association for Computing Machinery.

Dorri, A., Kanhere, S. S., and Jurdak, R. (2018). Multi-

agent systems: A survey. IEEE Access, 6:28573–

28593.

Giraldo, J., Sarkar, E., Cardenas, A. A., Maniatakos, M.,

and Kantarcioglu, M. (2017). Security and privacy in

cyber-physical systems: A survey of surveys. IEEE

Design & Test, 34(4):7–17.

Hafsi, K., Genon-Catalot, D., Thiriet, J.-M., and Lefevre,

O. (2021). Dc building management system with ieee

802.3 bt standard. In 2021 High Performance Switch-

ing and Routing (HPSR), pages 1–8. IEEE.

Institute, M. G. (2018). Notes from the ai frontier: Model-

ing the impact of ai on the world economy.

Kriaa, S., Pietre-Cambacedes, L., Bouissou, M., and Hal-

gand, Y. (2015). A survey of approaches combining

safety and security for industrial control systems. Re-

liability Engineering & System Safety, 139:156–178.

Miranda, J., M

akitalo, N., Garcia-Alonso, J., Berrocal, J.,

Mikkonen, T., Canal, C., and Murillo, J. M. (2015).

From the internet of things to the internet of people.

IEEE Internet Computing, 19(2):40–47.

Priyadarshini, Greiner, S., Massierer, M., and Aktouf, O.

(2023). Feature-based software architecture analysis

to identify safety and security interactions. In ICSA

2023, March 13-17, 2023, pages 12–22. IEEE.

Scherr, S. and Brunet, A. (2017). Differential in-

ﬂuences of depression and personality traits on

the use of facebook. Social Media + Society,

3(1):2056305117698495.

Song, H., Fink, G. A., and Jeschke, S. (2017). Security

and privacy in cyber-physical systems: foundations,

principles, and applications. John Wiley & Sons.

Thaci, A., Thaci, S., Zylbeari, A. K., and Baftijari, A. Y.

(2024). Economic impact of artiﬁcial intelligence.

KNOWLEDGE-International Journal, 62(1):21–25.

Thomas, D., Hristo, A., Patrick, M., G

obel, J. C., and Sven,

F. (2019). A holistic system lifecycle engineering ap-

proach – closing the loop between system architecture

and digital twins. Procedia CIRP, 84:538–544. CIRP

Design Conference 2019, Portgal.

to the European Parliament, C. (2019). Communication

from the commission to the european parliament, the

european council, the council the european economic

and social committee and the committee of the re-

gions, pp. 24, 2019 (climate change mitigation.

Van Looy, A. (2021). A quantitative and qualitative study

of the link between business process management

and digital innovation. Information & Management,

58(2):103413.

Winﬁeld, A., Michael, K., Pitt, J., and Evers, V. (2019).

Machine ethics: The design and governance of ethical

ai and autonomous systems. Proceedings of the IEEE,

107:509–517.

Xia, Y., Qin, T., Ding, W., Li, H., Zhang, X., Yu, N., and

Liu, T.-Y. (2017). Finite budget analysis of multi-

armed bandit problems. Neurocomputing, 258:13–29.

Human-Centric Dev-X-Ops Process for Trustworthiness in AI-Based Systems

295