DIN SPEC 92001-2

Artificial Intelligence – Life Cycle Processes

and Quality Requirements – Part 2: Robustness

Robust AI
CASE STUDY
DIN SPEC 92001-2

The background organizations must develop general requirements

The term Artificial Intelligence (AI) covers a num- for their specific AI risks. This is followed by a
ber of different approaches to optimize IT systems comprehensive AI risk assessment, consisting of
so that they can solve very specialized problems. the steps “Threat Model Analysis”, “Impact Analy-
Most AI algorithms are based on neural networks sis”, and “Robustness Evaluation”. Based on the
that benefit from the huge amounts of data avail- empirical evaluation of the robustness, a targeted
able today. These neural networks are loosely risk reduction is then carried out - for example,
modeled on the brains of living organisms: They through measures to defend against attacks or to
can process information independently, react to it strengthen the robustness. Depending on the real
and learn to solve problems themselves. H ­ owever, development and deployment context of the AI,
unlike the human brain they are susceptible to these steps should take place in digital, simulated,
even minor disturbances in the input data. As a and/or physical AI environments. The DIN SPEC
result, machine learning may no longer be able to also suggests that developers and operators
solve a problem in a sufficiently robust manner, divide their AI modules into different risk catego-
i.e. it may not be stable. However, because AI ries in order to build up their robustness in a
modules are increasingly used in highly sensitive risk-oriented way. DIN SPEC 92001-2 comprises
industries and business areas, robustness and 53 technical requirements, which are classified in
quality assurance throughout their entire life priority categories.
cycle are necessary.
The benefit
The DIN SPEC The DIN SPEC 92001-2 structures the very lively
The DIN SPEC 92001-2 is based on DIN SPEC research field of AI robustness and creates a
92001-1. The latter describes a quality metamodel detailed technical basis for robust and trust-
for AI modules, which introduces different phases worthy AI applications. It enables both developers
in their life cycle and defines robustness, func- and users of AI to manage AI risks in a modern
tionality and performance as well as comprehen- way by means of clearly formulated and pragmati-
sibility as fundamental quality pillars. Part 2 now cally balanced guidelines and recommendations
takes a closer look at the aspect of robustness for action. To this end, the risks associated with
under two specific angles. On the one hand, it the use of AI systems are analyzed and considered
deals with the robustness against mathematically as they occur in the various processes and phases
optimized disturbances, which are used to gener- in the life cycle of an AI module. “With the help of
ate malfunctions of AI, for example in the case of a DIN SPEC 92001-2, companies and other organiza-
targeted attack. On the other hand, robustness tions can make their AI-based software more
against naturally occurring signal disturbances or resistant to all conceivable disruptive processes,”
other impairments of data quality is necessary, explains Stephan Hinze, Managing Director of
which often occur when using AI. The DIN SPEC neurocat GmbH and initiator of DIN-SPEC 92001-2,
92001-2 structures extensive technical require- “It is a necessary building block in the AI strategy
ments along a systematic risk management pro- of every organization because it makes AI quality
cess. In the first step, “Scope, Context, Criteria”, assurance transparent and comprehensible.”

www.din.de/go/spec-en
 “DIN SPEC 92001-2 creates a
detailed technical basis for robust
and trustworthy AI applications.”

The collaboration
CASE STUDY
DIN SPEC 92001-2

About DIN SPEC

DIN SPEC 92001-2 was developed within 18 The success of an idea is often determined by how
months using the PAS (Publicly Available Specifi- quickly it is spread in the market. With the DIN
cation) procedure and is published in English. The SPEC, companies - from start-ups to medium-sized
project involved Acsioma GmbH, DFKI GmbH, and large enterprises - set standards within a few
EMEIA-GSA Automation, Ernst & Young AG, months in an agile and uncomplicated manner. The
Fraunhofer - Institute for Open Communication DIN SPEC is firmly linked to the names of the inno-
Systems FOKUS, Fraunhofer - Institute for Molec- vators, making it an effective marketing instrument
ular Biology and Applied Ecology (IME), GESTALT that leads to great acceptance among customers
Robotics GmbH, University of Applied Sciences and partners thanks to the recognized DIN brand.
(HAW), University of Applied Sciences Berlin DIN itself ensures that the DIN SPEC does not col-
(HTW) FB4 Economics, IABG mbH, Micropsi lide with existing standards and publishes them
industries GmbH, Microsoft Deutschland GmbH, internationally. A DIN SPEC can also be the basis for
neurocat GmbH, Otto-von-Guericke University a future DIN standard.
Magdeburg, Institute III: Philosophy, Robert
Bosch GmbH, Stiftung neue Verantwortung e.V., Five reasons for DIN SPEC
STILL GmbH, TÜV Süd Auto Service GmbH, Uni- → Fast: DIN SPEC can be created and
versity of Osnabrück and the University of published within a few months.
­Tübingen. → Worldwide recognition: Well established interna-
tionally, the DIN brand ensures maximum trust
The DIN SPEC 92001-2 is available for free in the market. Innovations and companies enjoy
­download under www.beuth.de high acceptance among users and investors.
→ Agile network: The DIN SPEC process promotes
exchange with relevant market participants.
This expands the network with key players:
Requirements of manufacturers and
customers are incorporated.
→ Easy handling: DIN organizes the entire DIN
SPEC project. This saves time, so you can
­concentrate on content and networking.
→ Direct plug & play: The DIN SPEC process
ensures that the innovation is aligned with the
current state of the art. Users can work with the
standard immediately and without hurdles.

www.din.de/go/spec-en DIN e. V. · Saatwinkler Damm 42/43 · 13627 Berlin

phone: +49 30 2601-2303 or -2607 · fax: +49 30 2601-42303
or -42607 · e-mail: [email protected]