ETSI TS 104 050 V1.1.1 (2025-03)

TECHNICAL SPECIFICATION
Securing Artificial Intelligence (SAI);
AI Threat Ontology and definitions

2 ETSI TS 104 050 V1.1.1 (2025-03)

Reference
RTS/SAI-005
Keywords
artificial intelligence
ETSI
650 Route des Lucioles
F-06921 Sophia Antipolis Cedex - FRANCE

Tel.: +33 4 92 94 42 00  Fax: +33 4 93 65 47 16

Siret N° 348 623 562 00017 - APE 7112B
Association à but non lucratif enregistrée à la
Sous-Préfecture de Grasse (06) N° w061004871

Important notice
The present document can be downloaded from the
ETSI Search & Browse Standards application.
The present document may be made available in electronic versions and/or in print. The content of any electronic and/or
print versions of the present document shall not be modified without the prior written authorization of ETSI. In case of any
existing or perceived difference in contents between such versions and/or in print, the prevailing version of an ETSI
deliverable is the one made publicly available in PDF format on ETSI deliver repository.
Users should be aware that the present document may be revised or have its status changed,
this information is available in the Milestones listing.
If you find errors in the present document, please send your comments to
the relevant service listed under Committee Support Staff.
If you find a security vulnerability in the present document, please report it through our
Coordinated Vulnerability Disclosure (CVD) program.
Notice of disclaimer & limitation of liability
The information provided in the present deliverable is directed solely to professionals who have the appropriate degree of
experience to understand and interpret its content in accordance with generally accepted engineering or
other professional standard and applicable regulations.
No recommendation as to products and services or vendors is made or should be implied.
No representation or warranty is made that this deliverable is technically accurate or sufficient or conforms to any law
and/or governmental rule and/or regulation and further, no representation or warranty is made of merchantability or fitness
for any particular purpose or against infringement of intellectual property rights.
In no event shall ETSI be held liable for loss of profits or any other incidental or consequential damages.

Any software contained in this deliverable is provided "AS IS" with no warranties, express or implied, including but not
limited to, the warranties of merchantability, fitness for a particular purpose and non-infringement of intellectual property
rights and ETSI shall not be held liable in any event for any damages whatsoever (including, without limitation, damages
for loss of profits, business interruption, loss of information, or any other pecuniary loss) arising out of or related to the use
of or inability to use the software.
Copyright Notification
No part may be reproduced or utilized in any form or by any means, electronic or mechanical, including photocopying and
microfilm except as authorized by written permission of ETSI.
The content of the PDF version shall not be modified without the written authorization of ETSI.
The copyright and the foregoing restriction extend to reproduction in all media.

© ETSI 2025.
All rights reserved.
ETSI
3 ETSI TS 104 050 V1.1.1 (2025-03)
Contents
Intellectual Property Rights . 4
Foreword . 4
Modal verbs terminology . 4
1 Scope . 5
2 References . 5
2.1 Normative references . 5
2.2 Informative references . 5
3 Definition of terms, symbols and abbreviations . 6
3.1 Terms . 6
3.2 Symbols . 7
3.3 Abbreviations . 7
4 From taxonomy to an ontology for secure AI . 7
4.1 Overview . 7
4.2 Formal expression of an ontology . 10
4.3 Relationship to other work . 11
5 Threat landscape . 13
5.1 Threat dimensions . 13
5.2 Attacks as instance of threat agent . 14
5.3 Adversarial Goals . 14
5.3.1 Violation of Confidentiality . 14
5.3.2 Violation of Integrity and Availability. 14
5.4 Threat modelling . 15
5.4.1 Attacker objectives . 15
5.4.2 Attack surface . 15
5.4.2.1 AI effect on impact and likelihood . 15
5.4.2.2 Data acquisition and curation . 16
5.4.2.3 Implementation . 16
5.4.2.4 Deployment . 16
5.4.2.5 Humans . 16
5.4.3 Trust model . 17
5.5 Statistics in AI and ML . 17
6 AI and SAI ontology . 18
6.1 Nouns, verbs, adverbs and adjectives . 18
6.2 Taxonomy and ontology . 18
6.3 Core SAI ontology relationships . 19
Annex A (informative): Cultural origins of ICT based intelligence . 22
Annex B (informative): Machine processing to simulate intelligence . 25
B.1 Overview of the machine intelligence continuum . 25
B.2 Expert systems . 25
B.3 Data mining and pattern extraction . 25
Annex C (informative): Bibliography . 26
C.1 AGI analysis . 26
C.2 AI in the context of threat analysis . 27
C.3 Societal and cultural references to AI . 27
History . 28

ETSI
4 ETSI TS 104 050 V1.1.1 (2025-03)
Intellectual Property Rights
Essential patents
IPRs essential or potentially essential to normative deliverables may have been declared to ETSI. The declarations
pertaining to these essential IPRs, if any, are publicly available for ETSI members and non-members, and can be
found in ETSI SR 000 314: "Intellectual Property Rights (IPRs); Essential, or potentially Essential, IPRs notified to
ETSI in respect of ETSI standards", which is available from the ETSI Secretariat. Latest updates are available on the
ETSI IPR online database.
Pursuant to the ETSI Directives including the ETSI IPR Policy, no investigation regarding the essentiality of IPRs,
including IPR searches, has been carried out by ETSI. No guarantee can be given as to the existence of other IPRs not
referenced in ETSI SR 000 314 (or the updates on the ETSI Web server) which are, or may be, or may become,
essential to the present document.
Trademarks
The present document may include trademarks and/or tradenames which are asserted and/or registered by their owners.
ETSI claims no ownership of these except for any which are indicated as being the property of ETSI, and conveys no
right to use or reproduce any trademark and/or tradename. Mention of those trademarks in the present document does
not constitute an endorsement by ETSI of products, services or organizations associated with those trademarks.
DECT™, PLUGTESTS™, UMTS™ and the ETSI logo are trademarks of ETSI registered for the benefit of its
Members. 3GPP™, LTE™ and 5G™ logo are trademarks of ETSI registered for the benefit of its Members and of the
3GPP Organizational Partners. oneM2M™ logo is a trademark of ETSI registered for the benefit of its Members and of ®
the oneM2M Partners. GSM and the GSM logo are trademarks registered and owned by the GSM Association.
Foreword
This Technical Specification (TS) has been produced by ETSI Technical Committee Securing Artificial Intelligence
(SAI).
NOTE: The present document updates and extends ETSI GR SAI 001 [i.20] prepared by ISG SAI.
Modal verbs terminology
In the present document "shall", "shall not", "should", "should not", "may", "need not", "will", "will not", "can" and
"cannot" are to be interpreted as described in clause 3.2 of the ETSI Drafting Rules (Verbal forms for the expression of
provisions).
"must" and "must not" are NOT allowed in ETSI deliverables except when used in direct citation.

ETSI
5 ETSI TS 104 050 V1.1.1 (2025-03)
1 Scope
The present document defines what an Artificial Intelligence (AI) threat is and defines how it can be distinguished from
any non-AI threat. The model of an AI threat is presented in the form of an ontology to give a view of the relationships
between actors representing threats, threat agents, assets and so forth and defines those terms (see also [1]). The
ontology in the present document extends from the base taxonomy of threats and threat agents described in ETSI
TS 102 165-1 [2] and addresses the overall problem statement for SAI presented in ETSI TR 104 221 [i.21] and the
mitigation strategies described in ETSI TR 104 222 [i.22]. Note that, although both technical reports are listed in
clause 2.2, they are indeed essential for understanding the scope of the present document.
NOTE 1: The ontology described in the present document applies to AI both as a threat agent and as an attack
target.
NOTE 2: The present document extends the content of ETSI GR SAI 001 [i.20], and retains significant elements of
its content where relevant for clarity.
2 References
2.1 Normative references
References are either specific (identified by date of publication and/or edition number or version number) or
non-specific. For specific references, only the cited version applies. For non-specific references, the latest version of the
referenced document (including any amendments) applies.
Referenced documents which are not found to be publicly available in the expected location might be found in the
ETSI docbox.
NOTE: While any hyperlinks included in this clause were valid at the time of publication, ETSI cannot guarantee
their long term validity.
The following referenced documents are necessary for the application of the present document.
[1] ISO/IEC 22989:2022: "Information technology - Artificial intelligence - Artificial intelligence
concepts and terminology".
NOTE: Many of the terms defined in the cited document above are also visible on the ISO Online Browsing
Platform: https://www.iso.org/obp/.
[2] ETSI TS 102 165-1: "Cyber Security (CYBER); Methods and protocols; Part 1: Method and
pro forma for Threat, Vulnerability, Risk Analysis (TVRA)".
2.2 Informative references
References are either specific (identified by date of publication and/or edition number or version number) or
non-specific. For specific references, only the cited version applies. For non-specific references, the latest version of the
referenced document (including any amendments) applies.
NOTE: While any hyperlinks included in this clause were valid at the time of publication, ETSI cannot guarantee
their long term validity.
The following referenced documents are not necessary for the application of the present document but they assist the
user with regard to a particular subject area.
[i.1] Alan Turing: "On computable numbers, with an application to the Entscheidungsproblem".
[i.2] Alan Turing: "Computing Machinery and Intelligence".
[i.3] Philip K. Dick: "Do androids dream of electric sheep?" (ISBN-13: 978-0575094185).
[i.4] Isaac Asimov: "I, robot" (ISBN-13: 978-0008279554).
ETSI
6 ETSI TS 104 050 V1.1.1 (2025-03) ®
[i.5] W3C Recommendation 11 December 2012: "OWL: OWL 2 Web Ontology Language
Document Overview (Second Edition)". ®
Working Group Note; 24 June 2014.
[i.6] RDF: RDF 1.1 Primer; W3C
[i.7] Cohen, Jacob (1960): "A coefficient of agreement for nominal scales". Educational and
Psychological Measurement. 20 (1): 37-46. doi:10.1177/001316446002000104. hdl:1942/28116.
S2CID 15926286. ®
[i.8] W3C Recommendation 16 July 2020: "JSON-LD 1.1: A JSON-based Serialization for Linked
Data".
[i.9] ETSI GS CIM 009 (V1.2.2): "Context Information Management (CIM); NGSI-LD API".
[i.10] "The Emergence Of Offensive AI".
[i.11] "Weaponizing Data Science for Social Engineering: Automated E2E Spear Phishing on Twitter".
[i.12] Li Chen, Chih-Yuan Yang, Anindya Paul, Ravi Sahita: "Towards resilient machine learning for
ransomware detection".
[i.13] Alejandro Correa Bahnsen, Ivan Torroledo, Luis David Camacho and Sergio Villegas:
"DeepPhish: Simulating Malicious AI".
[i.14] Common Weakness Enumeration Project.
[i.15] ETSI TS 118 112: "oneM2M; Base Ontology".
[i.16] The Smart Appliances REFerence (SAREF) ontology.
[i.17] ETSI TS 102 165-2: "CYBER; Methods and protocols; Part 2: Protocol Framework Definition;
Security Counter Measures".
[i.18] ETSI TR 104 048: "Securing Artificial Intelligence (SAI); Data Supply Chain Security".
[i.19] Andrew Marshall, Jugal Parikh, Emre Kiciman and Ram Shankar Siva Kumar: "Threat Modeling
AI/ML Systems and Dependencies".
[i.20] ETSI GR SAI 001: "Securing Artificial Intelligence (SAI); AI Threat Ontology".
[i.21] ETSI TR 104 221: "Securing Artificial Intelligence (SAI); Problem Statement".
[i.22] ETSI TR 104 222: "Securing Artificial Intelligence; Mitigation Strategy Report".
3 Definition of terms, symbols and abbreviations
3.1 Terms
For the purposes of the present document, the terms given in ETSI TR 104 221 [i.21], ISO/IEC 22989 [1] and the
following apply:
Artificial General Intelligence (AGI): applying intelligence to any intellectual task, at a level equivalent to a human
NOTE: AGI is also termed Strong AI.
Artificial Intelligence (AI): ability of a system to handle representations, both explicit and implicit, and procedures to
perform tasks that would be considered intelligent if performed by a human
NOTE: From ETSI TR 104 221 [i.21].
Artificial Narrow Intelligence (ANI): applying intelligence to only one context
EXAMPLE: Autonomous driving, speech recognition.
ETSI
7 ETSI TS 104 050 V1.1.1 (2025-03)
NOTE: ANI is also termed Weak AI.
Artificial Super Intelligence (ASI): extending beyond AGI to apply intelligence to a level significantly beyond those
of humans across a comprehensive range of categories and fields of endeavour
cognition: mental action or process of acquiring knowledge and understanding through thought, experience, and the
senses
predicate: part of a sentence or clause containing a verb and stating something about the subject
NOTE: In the context of the present document as applied to RDF statements the predicate illustrates the nature of
the relationship between two objects or concepts.
reasoning: application of learned strategies in order to solve puzzles, and make judgments where there is uncertainty in
either the input or the expected outcome
3.2 Symbols
Void.
3.3 Abbreviations
For the purposes of the present document, the abbreviations given in ETSI TR 104 221 [i.21] and the following apply:
AGI Artificial General Intelligence
AI Artificial Intelligence
ANI Artificial Narrow Intelligence
ASI Artificial Super Intelligence
CAV Connected and Autonomous Vehicles
CIA Confidentiality Integrity Availability
CVE Common Vulnerabilities and Exposures
CVSS Common Vulnerability Scoring System
CWE Common Weakness Enumeration
GAN Generative Adversarial Networks
ICT Information Communications Technology
IQ Intelligence Quotient
IT Information Technology
ITS Intelligent Transport Systems
JSON JavaScript Object Notation
LD Linked Data
ML Machine Learning
NGSI Next Generation
NGSI-LD Next Generation Service Interface - Linked Data
OWL Ontology Web Language
RDF Resource Description Framework
RL Reinforcement Learning
SAI Securing Artificial Intelligence
TVRA Threat Vulnerability Risk Analysis
UML Unified Modelling Language
XSS Cross Site Scripting
4 From taxonomy to an ontology for secure AI
4.1 Overview
An ontology in information science identifies a set of concepts and categories within a particular field of knowledge
that shows the properties of the concepts and categories and the relations between them.
ETSI
8 ETSI TS 104 050 V1.1.1 (2025-03)
This overview illustrates and demonstrates how the various concepts that are taken for granted in the security standards
space are implicit as taxonomies. The overview extends to illustrate that by adopting a broader understanding of these
implicit taxonomies in the form of an ontology, in which concepts are related, will help in making systems more
resilient against AI attackers, or which make better use of AI in defence.
NOTE 1: The model of ontology from philosophy is the study of being, and addresses concepts such as becoming,
existence and reality. For many, the ultimate aim of AI is general intelligence i.e. the ability of a single
machine agent able to learn or understand any task, covering the range of human cognition. If and when
AI moves closer to any concept of independent sentience, there will be increasing overlap between the
worlds of information science and philosophy. However, this is likely to be decades away at least, and so
the present document focusses on so-called weak AI: the use of software to perform specific, pre-defined,
reasoning tasks. Also, in the philosophical domain there is a degree of crossover in the role of intelligence
and the role of ethics. The present document does not attempt to define the role of ethics other than to
reflect that in an ontology of intelligence that there are various schools of ethics that apply. So, an
intelligence framework is influenced by its ethical framework, where the impact of the ethical framework
can be realized in various ways.
In many domains that apply some form of AI, the core data model is presented in an ontological form and from that it is
possible to apply more sophisticated search algorithms to allow for semantic reasoning. The technical presentation of an
ontology is therefore significant of itself as it can pre-determine the way in which the programming logic is able to
express intelligence. Ontologies, in the context of a semantic web, are often designed for re-use. In addition to
conventional ontologies and the use of Resource Description Framework (RDF) [i.6] notations, there is growth in the
use of Linked Data extensions to data passing mechanisms used widely on the internet.
EXAMPLE 1: JSON-LD [i.8] has been designed around the concept of a "context" to provide additional
mappings from JSON to an RDF model. The context links object properties in a JSON document
to concepts in an ontology.
EXAMPLE 2: NGSI-LD [i.9]. The term NGSI (Next Generation Service Interfaces) was first developed in work
by the Open Mobile Alliance and has been extended using concepts of Linked Data to allow for
wider adoption of ontologies and semantic as well as contextual information in data-driven
systems.
As a pre-cursor to the development of a threat ontology for AI based threats, there are a number of threat taxonomies,
some found in ETSI TS 102 165-1 [2] and in ETSI TS 102 165-2 [i.17]. These can serve as a starting point for the
definition of a threat ontology, and more specifically of an AI threat ontology.
ETSI
9 ETSI TS 104 050 V1.1.1 (2025-03)
cd ThreatTree
Threat
Interception Manipulation Repudiation DenialOfServ ice
Forgery
UnauthorisedAccess Masquerade InformationCorruption InfornationLoss

Figure 1: Threat tree (from ETSI TS 102 165-1 [2]) as a taxonomy
In the conventional taxonomy, as in Figure 1 for threats, the core relationship between entities is of type "is a", thus
Forgery "is a" Manipulation, "is a" Threat. The relationships in a conventional taxonomy are often unidirectional,
whereas in an ontology the normal expectation is that relationships are bidirectional and asymmetric.
EXAMPLE 3: Trust is asymmetric, a pupil is expected to trust a teacher, whereas the teacher is not expected to
trust the child.
A simple taxonomy such as in Figure 1 does not easily express side channel attacks, or composite attacks, nor does it
capture the asymmetric relationships of things like trust.
EXAMPLE 4: In order to perform a masquerade attack it is often necessary to first have intercepted data, or in
order to corrupt data it can be necessary to first have masqueraded as an authorized entity.
Many of the forms of attack on AI that are described in the SAI Problem Statement (ETSI TR 104 221 [i.21]) are in the
manipulation tree: data poisoning is a form of information corruption; incomplete data is a form of information loss.
The relationship in these cases are "modifies", and "is modified by". Similarly, the terms "threat" and "vulnerability" as
defined in ETSI TS 102 165-1 [2] are loosely expressed in the form of ontological relationships. Thus, threat is defined
as the potential cause of an incident that may result in harm to a system or organization, where a threat consists of an
asset, a threat agent and an adverse action of that threat agent on that asset, and a threat is enacted by a threat agent, and
may lead to an unwanted incident breaking certain pre-defined security objectives.
NOTE 2: The nature of data poisoning is complex to clearly identify. The consequence of data poisoning are to
limit the ability of the reasoning element of AI to reason towards the "right" solution, but rather to lead
the reasoning algorithms to an invalid answer, often in favour of the attacker.
NOTE 3: The deliberate introduction of poisoned data may lead to the AI system exhibiting bias, where the original
(non-poisoned) data may have zero biases.
NOTE 4: Whilst it is suggested that incomplete data is a form of information loss the attack vector in an AI system
may be quite different than that from non-AI systems.
ETSI
10 ETSI TS 104 050 V1.1.1 (2025-03)
The structure of the term vulnerability has a similar ontological grouping of relationships, being modelled as the
combination of a weakness that can be exploited by one or more threats. A more in-depth examination of the problems
of and from AI is found in the SAI Problem Statement [i.21], and in the SAI report on mitigation strategies [i.22].
4.2 Formal expression of an ontology
There are many ways to express an ontology in information science. The most common are:
• OWL - Ontology Web Language [i.5]
• RDF - Resource Description Framework [i.6]
It should be noted, however, that OWL and RDF, whilst common when referring to ontologies, are not equivalent but
are mutually supportive.
A simple model that underpins both OWL and RDF is the subject-predicate-object grammar structure (see Figure 2).
However, there is also a more complex set of data structures that also look like the object-oriented design concepts (e.g.
inheritance, overloading) underpinning design languages such as UML, and coding languages such as C++, Swift and
Java. Such taxonomical classifications are also common in science, particularly in the biological sciences.

Figure 2: Simplified model of grammar underpinning Ontology
An ontology is expected to consist of the following elements:
• Classes, also known as type, sort or category.
• Attributes, which describe object instances, such as "has name", "has colour", "by definition has a".
EXAMPLE 1: A protected object belongs to class network object, of sub-type router, with name "Router-1" and,
by definition, has 1 or more Ethernet ports.
EXAMPLE 2: Ransomware belongs to class threat, of subclass denial-of-service, with attribute file-encryption.
• Relationships, as outlined in 4.1 identifies how one class is associated to another.
Expanding from the taxonomy in [i.5], threat is modelled as one class, with threat agent modelled as another. This is
then consistent with the definitions given for the terms "threat" and "vulnerability", and for the relationship to assets as
the subject or object in the simplified grammar of ontology.
In the gap between an ontology and natural language, the present document classifies concepts around intelligence as
nouns, and relationships as verbs, adverbs, adjectives.
NOTE: Whilst there is a risk in trying to explain AI only by mapping to programming constructs (e.g. objects and
classes), or only from data modelling (e.g. tables, lists, numbers, strings and the relationships or type
constraints a data model can impose) it is not addressed by the present document but is considered in
ETSI TS 102 165-1 [2].
As stated above, an ontology is often described as a specification of a conceptualization of a domain. The result of such
an approach to an ontology is to provide standardized definitions for the concepts of a specific domain. In the structure
of a technical standard the ontology defines classes (concepts) for sets of objects in the domain that have common
characteristics. The objects include specific events, actions, procedures, ideas, and so forth in addition to physical
objects. In addition to the concepts, the ontology describes their characteristics or attributes, and defines typed
relationships that may hold between actual objects that belong to one or more concepts.
ETSI
11 ETSI TS 104 050 V1.1.1 (2025-03)
Information in an ontology is conventionally encoded in languages such as RDF and in representations such as OWL, as
a list of triplets (the "subject-predicate-object" concept), where the subject is the domain under analysis, the objects are
all relevant concepts affecting the subject and the predicate defines the relationship between them.
For illustrative purposes, this can be expressed as a mathematical representation of a linear system:
� � �0 � �1�1 � �2 �2 � �3 �3 � … � �� �� � µ
�� � 0
where Y is the domain variable to be explained by the ontology, X are the concepts as explicative variables, and
coefficient β represents the relationship of the explicative variables over the variable Y, and µ is the error factor for the
"unknown" concepts in Y.
In regard to standardization, ontologies, when formally modelled, provide explicit knowledge models for particular
domains that can assist in both structuring the problem and in identifying where standards can assist in specifying the
nature of the domain in such a way that it becomes known.
4.3 Relationship to other work
In the scope of the present document an ontology is also developed to assist in the development of strategies in securing
AI. This addresses the modes in which AI can exist in a system, shown figuratively in Figure 3 below.

Figure 3: Modes of application of AI in networks and services
AI can be deployed in attack mode against components or systems, defence mode in countering attacks on components
or systems, and proactively in understanding attacks on components and systems. Underpinning both attack and defence
modes is the goal of understanding the problem associated to AI and the risk to the system of AI.
ETSI
12 ETSI TS 104 050 V1.1.1 (2025-03)
In ETSI TR 104 221 [i.21], "SAI Problem Statement", the following definition of AI is offered:
Artificial intelligence is the ability of a system to handle representations, both explicit and implicit, and
procedures to perform tasks that would be considered intelligent if performed by a human.
In ETSI TR 104 221 [i.21], the intent is to describe the challenges of securing AI-based systems and solutions, in an
environment where data, algorithms and models (in both training and implementation environments) are portrayed as
challenges to the overall systems analysis and understanding in respect of AI as a system threat. In [i.21] and in
ETSI TR 104 222 [i.22] there is a subtle change of emphasis with regards to data that is input to a system where the
dominant AI mechanism is machine learning. In the specific subset of AI that is Machine Learning (ML), there are
further modes of learning that can be defined. As noted above, the role of ontologies is implied in most ML systems as a
means of structuring the input. In practice most ontologies are incomplete: they tend to be domain specific, whereas in
practical systems an entity can exist in more than one domain.
EXAMPLE: A potato will naturally exist in an ontology describing tubers and variants of the nightshades, but
will also naturally exist in an ontology describing foodstuffs and diet. Domain specific knowledge,
knowing that a potato is related to a tomato in the family of nightshades, would not necessarily
reveal the existence of poutine. Or in other words there may not be an obvious link between
domains.
Figure 4: Typical machine learning lifecycle from ETSI TR 104 221 [i.21]
The ML lifecycle considered in ETSI TR 104 221 [i.21] identifies a number of ML strategies:
• Supervised learning - where all the training data is labelled, and the model can be trained to predict the output
based on a new set of inputs.
• Semi-supervised learning - where the data set is partially labelled. In this case, even the unlabelled data can be
used to improve the quality of the model.
• Unsupervised learning - where the data set is unlabelled, and the model looks for structure in the data,
including grouping and clustering.
• Reinforcement learning - where a policy defining how to act is learned by agents through experience to
maximize their reward; and agents gain experience by interacting in an environment through state transitions.
In each case ML can be used for both classification problems, and for prediction problems.
ETSI
13 ETSI TS 104 050 V1.1.1 (2025-03)
In applying annotations or labels to data, it is common practice to first define the domain data using an ontology, in its
simplest form of a semantic data set. As an ontological or semantic data definition is unlikely to be complete, one role
often assigned to AI/ML is to further develop the data description, by identifying additional patterns through finding
new correlations and asserting new causations.
Attack strategies against the ML workflow (illustrated in Figure 5) may apply to each stage or process flow, and the risk
associated to each is assessed independently. The role of risk management is covered in more detail in clause 5 of the
present document.
Figure 5: Machine learning model workflow from ETSI TR 104 222 [i.22]
5 Threat landscape
5.1 Threat dimensions
The TVRA model in ETSI TS 102 165-1 [2] states that "A threat agent enacts a specific attack against a system
weakness to exploit a vulnerability". AI shall be considered in the context of each of the items in this statement
identified with bold text, in both offensive and defensive contexts. The SAI problem statement in ETSI
TR 104 221 [i.21] identifies ways in which a threat agent can invoke particular forms of attack, while in ETSI
TR 104 222 [i.22] a number of specific mitigations are identified. In each of [i.21] and [1] the focus is on attack and
defence of AI-inspired attacks on AI systems, whereas in the present document the focus is on the understanding of
what AI means and of defining the AI domain itself.
Acquired intelligence, i.e. intelligence from learning, requires knowledge of data semantics (i.e. what data elements
mean) and data context (i.e. how data elements are related), and conventional domain ontologies offer this form of data
labelling. The richer the ontology of the input data, i.e. the more that data is labelled, the closer the ontology is to the
world model required by the AI to represent the world view for the intelligence in the machine. In other words,
semantic labelling is a major step forward in gaining an understanding of data.
EXAMPLE: The numerical value 42 can be syntactically represented as a signed integer which in computing
terms means certain functions can be applied to it (arithmetic functions say) and a compiler will be
able to warn if functions will fail based on knowledge of the syntax. However, of itself the value
of the integer does not confer knowledge whereas adding a semantic label to it allows reasoning to
be applied. Simple semantic labels can be seen in the names given by programmers to constants
and variables, but in the wider context semantics have to be transferred with the data in order that
the receiver has knowledge of what the value means, or is associated to.
ETSI
14 ETSI TS 104 050 V1.1.1 (2025-03)
5.2 Attacks as instance of threat agent
According to a 2019 report by Forrester, 86 % of cybersecurity decision makers are concerned about the offensive use
of AI by threat actors [i.10]. As with many other organizations, adversaries are increasingly looking to AI to automate,
scale and speed up activities which are currently conducted manually. This is particularly where malicious actors target
indiscriminately, where the lower likelihood of a successful single attack is offset by the opportunity to attack many
more targets. As such, although the impact of individual threats is unlikely to change, the scale of attacks, the likelihood
of attacks being successfully carried out, and the difficulty in responding or remediating in a timely manner may all
increase dramatically.
Opportunities range across the phases of the attack chain, from reconnaissance to exfiltration and impact. Public data,
particularly social media, presents a significant opportunity for AI-based exploitation. For example, supervised and
unsupervised learning can be used as tools to mine data for large-scale target discovery and spearphishing email
generation, and reinforcement learning for conducting automated phishing [i.11].
AI can be used to evade defence mechanisms, including those defences based themselves on AI techniques:
EXAMPLE 1: Generative Adversarial Networks (GANs) can be used to evade ransomware detection [i.12].
EXAMPLE 2: The use of deep neural networks may be used to evade phishing detection [i.13].
AI tools can also be used to enable other kinds of attack. ML approaches are increasingly being demonstrated in side-
channel analysis as an alternative to traditional statistical tools. The increasing sophistication of AI-based techniques for
generating fake biometric data, and creating falsified images (so-called DeepFakes), can also pose a threat to, and
undermine, trust in the integrity and authenticity of data in all aspects of a business:
EXAMPLE 3: In one case, AI-based voice-mimicking software was used to defraud an energy company of
$240 000, by convincing an employee to make a fraudulent bank transfer.
EXAMPLE 4: Deepfakes, a term applied to synthetic media in which a person in an existing image or video is
replaced with someone else's likeness, has been used in a number of situations both positive and
negative, including replacement of images to commit fraud or misdirection, and in more positive
environments to place a dead actor in a film.
In all these examples, the use of AI is unlikely to change the impact of a successful exploitation. However, it can
increase the likelihood of an organization being targeted and/or of attack attempts being successful, and hence can
increase the overall risk.
5.3 Adversarial Goals
5.3.1 Violation of Confidentiality
As mentioned above, data is a crucial asset in an AI-based system. By definition, information about training data is
encoded in a model itself: a model can be considered the aggregated understanding of a scenario or task derived from
analysis of many examples of that scenario. Techniques exist whereby an adversary can infer aspects of this
information, for example reconstructing training data examples (so-called model inversion) which represents a violation
of confidentiality, and potentially privacy where a model has been trained on personal data. Similarly, interrogation of a
model can leak information about the model itself, which can represent a leak of proprietary information. This can be
particularly damaging where a business model is based around a well-trained model.
5.3.2 Violation of Integrity and Availability
As described in clause 5, the role of AI in a system is usually to make inferences about data to enable downstream
decision making (human or machine), or to carry out actions based on input data. Compromise of the AI component can
hence lead to a violation of integrity of the system, in that inferences and decisions will be inaccurate, and overall
system performance will be degraded. If significant enough, this degradation can constitute a loss of availability, either
of the component itself or of the whole system, depending on the AI's role in that system. The various modes of
compromise are described in ETSI TR 104 221 [i.21].
ETSI
15 ETSI TS 104 050 V1.1.1 (2025-03)
One variant of availability compromise that particularly applies in an AI context is "reputational compromise". AI is not
well understood in many domains, and one of the purposes of the present document is to offer a wider understanding of
AI. Users may be wary of trusting model outputs, especially when model reasoning is difficult to explain or is
sufficiently different from human reasoning. An attacker can target an AI system in an attempt to damage trust in AI
itself and disrupt or damage an organization by preventing or reversing the adoption of AI technologies.
A key aspect of the integrity definition of the CIA model is the ability to reverse the effects of attacks. The relationship
between data and a model, as well as the probabilistic nature of models themselves, make understanding and reversing
adversarial action more challenging for AI systems than for traditional software. Once a model is trained, it is extremely
difficult to undo the effect of malicious datapoints without significant retraining, which requires large amounts of
reliable training data and compute. Techniques exist to harden some types of AI against adversarial input, either as a
preventative or remediation measure; these are areas of active research.
5.4 Threat modelling
5.4.1 Attacker objectives
As with any cyberattack, an adversary will ultimately be aiming to extract information from a system or affect its
operation in some way. The adversary can choose to do so using AI, or by attacking AI components, but ultimately the
objective will be to affect a system or the information within it. Where an AI component of a system is used to provide
context for decision making within a system, or make decisions itself, then compromise of the AI component will affect
those decisions. The effect of compromising decisions will depend on the design and purpose of the system.
On a more granular level, attacks on AI systems can be thought of as aiming to force a model to do, learn or reveal the
wrong things:
• Do - the actor aims to engineer an input to a model such that the output will be incorrect. The actor has control
over the input but not the model itself. This class of attack is known as evasion. An example would be a
malware author manipulating an executable binary so that an ML-based security product classifies that binary
as benign software.
• Learn - the actor wishes to poison a model such that it will fail to operate as the intended, in a targeted or
indiscriminate way. The actor has control over the data and/or model. The actor may be looking to degrade the
overall performance of a model (functionally a denial-of-service attack), or to introduce a backdoor or trojan.
In the former case, the degradation or disruption can be the actor's ultimate aim, or they wish to us
...