Article 15 · Chapter III: High-Risk AI Systems
Accuracy, Robustness and Cybersecurity
High risk
Summary
High-risk AI systems must achieve, throughout their lifecycle, an appropriate level of accuracy, robustness and cybersecurity. Providers must declare the relevant performance metrics in instructions for use; resilience must include defences against adversarial examples, data poisoning, model evasion, and confidentiality attacks.
Key obligations
- Declare accuracy and robustness metrics in instructions for use
- Implement technical redundancy (fail-safes, backups)
- Protect against ML-specific threats: poisoning, evasion, model inversion