AR-004HochAnbieterDetektiv
Tests auf adversarielle Robustheit
Sofern für den Anwendungsfall des KI-Systems relevant, müssen Anbieter die Robustheit gegenüber adversariellen Eingaben testen, die darauf ausgelegt sind, Fehlklassifikationen, Datenvergiftung während des Trainings oder andere sicherheitsrelevante Ausfallmodi zu verursachen.
Artikel:Article 15(5)Article 15(4)
Nachweisbeispiele
- Adversarial test suite description
- Red-team exercise report
- Adversarial robustness metrics
Normen
ISO 42001:2023 §8.4NIST AI RMF
Verwandte Kontrollen
- AR-001Accuracy Level Definition and CommitmentArticle 15(1) requires that high-risk AI systems are designed and developed to achieve an appropriate level of accuracy, robustness, and cybersecurity,…
- AR-002Accuracy Measurement and VerificationProviders must measure and verify the accuracy of the AI system against defined metrics using representative test sets, documenting the methodology,…
- AR-003Robustness Testing Against Input VariationsProviders must test the AI system's robustness against input variations that may arise in real-world deployment, including incomplete data, noisy inputs,…
- AR-005Cybersecurity Measures for AI SystemsArticle 15(5) requires that high-risk AI systems are resilient against unauthorised third-party attempts to alter their outputs, and providers must…
- AR-006Fallback Procedures for System FailuresProviders must design and implement fallback procedures that activate when the AI system cannot operate within its designed parameters, ensuring safe…
- AR-007Input Validation and Data Quality ChecksProviders must implement input validation mechanisms that verify the quality, format, and plausibility of data before it is processed by the AI system,…
Verwandte KI-VO-Begriffe
- AccuracyThe requirement that high-risk AI systems achieve an appropriate level of accuracy in relation to their intended purpose, as specified in the technical documentation. Providers must declare the level of accuracy in the instructions for use.
- RobustnessThe ability of a high-risk AI system to maintain its level of performance under adverse conditions — including technical limitations, adversarial inputs, errors, or unexpected situations — or within foreseeable operating conditions outside the intended purpose.
- CybersecurityThe requirement that high-risk AI systems are resilient against attempts by third parties to alter their use, behaviour, or performance in ways that could result in risks to health, safety, or fundamental rights, including protection against data poisoning, adversarial examples, and model evasion attacks.
Erweitern Sie, wenn Nachweise rechtskräftig sein müssen.
Der kostenlose Fragebogen liefert erste Signale. Die Vollständige Bewertung verwandelt reale Systemdaten in einen auditierten Entscheidungsnachweis: Faktenextraktion, Komponententrennung, Nachweise, nationale Vorgaben und ein prüffähiges Dossier.
Kostenlose Risikovorschau starten