Part 1, the overview of robustness assessment
Part 1 gives background on methods already used to judge how robust a neural network is. It treats robustness as how well an AI system holds its performance whatever happens, with input conditions as its main focus.
The requirement in plain English
Part 1 gives background on methods already used to judge how robust a neural network is. It treats robustness as how well an AI system holds its performance whatever happens, with input conditions as its main focus.
What the public record says
The public view gives the scope, the terms and the introduction of Part 1. It defines neural network, input data, test data, validation data, robustness, testing, validation and verification.
Why it matters to you
Part 1 is the vocabulary the series assumes. Its definition of robustness is broader than the FDIS stability term: any circumstances, not specified perturbations, and level of performance rather than invariant output. The two are not in conflict, because they answer different questions, but a claim that uses one word for both will confuse a reviewer. Part 1 also gives the testing vocabulary that a statistical plan needs: test data, validation data, input data and the difference between validation and verification.
Evidence a reviewer asks for
- The glossary row for robustness, with Part 1 cited as its source.
- A note saying which definition a claim uses: the Part 1 robustness term or the FDIS stability term.
- The system boundary and the performance measure used in any robustness statement.
Checks to run this month
- Does the claim use robustness and stability as if they were the same?
- Is the performance measure named?
- Is the definition cited to Part 1 or to the FDIS?
Dates
Part 1 published as a Technical Report in 2021. FDIS registered 2026-08-21.
Cross-references
- Clause 3.1, Stability
- Robustness as a reliability sub-characteristic
- ISO/IEC 22989:2022, the vocabulary
What this page does not state
Part 1's methods chapters are not in the public view.