DIN 92001 defines a quality metamodel over the lifecycle of AI and identifies performance, stability, functionality, and traceability of AI as the overarching goals of AI quality assurance.
DIN SPEC 92001-2 has been publicly available since December 2020. The document of more than 40 pages is extensive and detailed. On this page you will find a short and abbreviated overview. You can download the entire DIN document unsing the following link .
Definition of robust AI
AI can be assumed to be robust if intentional interference by harmful adversaries and by unintentional interference due to changes in the input data do not impose risk on AI stakeholders. Both disturbances represent a safety and a protection problem.
AI must cope with faulty, noisy, unknown or
constructed input data. In complex environments, robustness is a key AI quality criterion. Even state-of-the-art AI modules are susceptible to these types of glitches. When companies want to make AI robust, there are two dimensions to focus on.
- intentional interference by harmful adversaries
- unintentional disturbance due to changes in data input
6 steps to improve your AI quality
How to implement a quality cycle for robust artificial intelligence based on DIN SPEC 92001-2?
- Match the tasks to be undertaken by the AI with your organization's risk appetite.
- Define "AI malfunction" per automated task
- Identify opponents, think about possible attacks and prepare defense strategies
- Monitor your input data and be ready to retrain your AI on new data
- Add error-causing data to your test data and calculate its impact after each training of the AI
- Document your strategy and repeat the above steps regularly
Already in November the Informatik Aktuell in a wide-ranging article about our KI-Dreieck®, which is based on DIN SPEC 92001. This, as well as our software Konfuzio, continuously expanded to exceed the latest demands on quality and robustness.