Skip to main content

EU AI Act Article 15: Accuracy and Robustness

What EU AI Act Article 15 requires for high-risk AI accuracy, robustness against adversarial attacks, fail-safe mechanisms, and cybersecurity.

Article 15: Accuracy, Robustness and Cybersecurity

High-risk AI systems must achieve appropriate levels of accuracy, robustness, and cybersecurity throughout their lifecycle.

Providers must declare accuracy metrics, build resilience against errors and adversarial attacks, and implement technical safeguards to prevent unauthorized manipulation of system behavior.

Core Requirements

Declared Accuracy Metrics

Define and document appropriate accuracy levels and metrics in the instructions for use. Accuracy must be measurable and verifiable throughout the system lifecycle.

Resilience to Errors and Faults

Systems must handle errors, faults, and inconsistencies gracefully. Implement redundancy mechanisms such as backup systems and fail-safe plans to maintain consistent performance.

Resilience to Adversarial Attacks

Protect against data poisoning and model poisoning attacks that corrupt training data or model parameters. Systems must detect and mitigate attempts to manipulate inputs or weights.

Protection Against Evasion and Confidentiality Attacks

Guard against model evasion attacks that exploit decision boundaries and confidentiality attacks that extract sensitive model information or training data.

Technical Cybersecurity Solutions

Deploy appropriate security measures to prevent and control adversarial attacks and model vulnerabilities. Solutions must address the specific risks of AI system exploitation.

Implementation Steps

Adversarial Testing Pipelines

Integrate adversarial robustness testing into CI/CD. Generate adversarial examples, test model boundaries, and validate defenses against known attack patterns.

Accuracy Benchmarking

Establish baseline accuracy metrics aligned with your declared levels. Track accuracy across model versions, datasets, and deployment environments to detect degradation.

Redundancy and Fail-Safe Design

Build fallback mechanisms for critical system paths. Implement backup models, safety monitors, and human-in-the-loop triggers when confidence drops below thresholds.

Model Supply Chain Security

Scan dependencies, base models, and training datasets for vulnerabilities. Verify the provenance of external models and validate integrity of training pipelines.

The Accuracy-Robustness Tradeoff

Optimizing for benchmark accuracy can reduce adversarial robustness. Providers must balance performance goals with security requirements and document design decisions that navigate this tradeoff.

Frequently Asked Questions

Related Articles


How Scanara Helps

Scanara automates EU AI Act compliance from code to dossier. Connect your GitHub repos and get compliance reports in minutes.