The results of artificial intelligence are not always reliable. This uncertainty can be measured to make the use of AI safer. A new DIN standard provides initial suggestions for this.

With the rapid progress of artificial intelligence (AI) came the big question of trust: How reliable are the results of AI? Where does the data for machine learning (ML) come from? Debates about fabricated citations and AI-generated photos that look deceptively real call for caution in handling technology. This is all the more true if it is part of safety-critical applications, for example in medicine or autonomous driving.

When an ML module is used in practice, it is inevitable that scenarios occur that have not been taken into account in the development of the technology. Because reality is complex – so complex that training and test data for machine learning only partially reflect it. By quantifying and estimating this systematic uncertainty, the use of ML modules within an overall system can be made safer and more robust.

Assistance for industry and science

This is where the new DIN SPEC 92005 comes in: Under the title "Artificial Intelligence – Quantification of Uncertainties in Machine Learning", the German Institute for Standardization (DIN) defines essential terms for so-called Uncertainty Quantification in Machine Learning. The document formulates requirements and guidelines and provides application examples. The aim is to give stakeholders in industry and science advice on how to deal with the quantification of uncertainties in practice.

"An important pillar for AI security"

"The quantification of uncertainty is a broad and intensively researched field. But so far there has been a lack of standardization that reflects the current state of the art," says Dr. Lukas Höhndorf from Industrieanlagen-Betriebsgesellschaft mbH, initiator of the new DIN SPEC. "Being able to estimate the uncertainty associated with models, algorithms and predicted results is an important pillar for the security of AI, especially in safety-critical systems. "

Many factors influence results

Not only incomplete or unclear data and scenarios can affect system environments with ML modules – also extreme weather conditions during autonomous driving or the unconventional behavior of other actors in the application environment can shape the results of AI.

"Artificial intelligence is the basis for many applications that cannot be realized with conventional software. In order for it to be used in sensitive application contexts, AI systems must be secure, robust and trustworthy. This DIN SPEC makes an important contribution to ensuring this," says Dr. Maximilian Porechkin, consortium leader of the CERTIFIED AI project and team leader AI Assurance and Certification at the Fraunhofer Institute for Intelligent Analysis and Information Systems IAIS.

The new DIN SPEC is a result of the CERTIFIED AI research project, in which DIN cooperates with the Fraunhofer Institute for Intelligent Analysis and Information Systems IAIS, the Federal Office for Information Security (BSI) and other scientific and industrial partners. Its aim is to ensure technical reliability and responsible use of technology. The project is funded by the Ministry of Economics, Industry, Climate Protection and Energy of the State of North Rhine-Westphalia. The standard was developed by a high-profile consortium of experts from the following organizations: Industrieanlagen-Betriebsgesellschaft mbH IABG, the Fraunhofer Institutes for Intelligent Analysis and Information Systems IAIS, for Cognitive Systems ICS as well as for Experimental Software Engineering IESE, ai.dopt, DEKRA Digital GmbH, e:fs TechHub GmbH, Giesecke+Devrient, Helsing GmbH, Leibniz University of Applied Sciences, Federal Network Agency for Electricity, Gas, Telecommunications, Post and Railways (BNetzA), Institute for Occupational Safety of the German Social Accident Insurance (IFA).