D2.6 ML Model Certification – v1

This deliverable presents the initial design, architecture, and implementation state of the machine learning (ML) model evidence extractors of WP2, which we call AI-SEC. They contribute to the key result KR1-EXTRACT of EMERALD, a framework to continuously extract knowledge from well-trained ML models and prepare suitable evidence based on them. EMERALD follows a knowledge graph-based approach to provide a unified view of the cloud service under certification at different layers of the service, ranging from the infrastructure layer (e.g., virtual resources), to the business layer (e.g., policies and procedures), to the implementation layer (e.g., source code files) and data layer (e.g., increasingly used AI models) in cloud applications. The ML model evidence extractors, developed in Task 2.4 and described in this deliverable, aim at identifying critical security-related features, such as adversarial robustness, privacy, security, and explainable AI. Other related deliverables in WP2, all due at project Month 12 (October 2024), provide functional and technical details on further evidence extractors from different sources, i.e., D2.2 on source code evidence extraction, D2.4 on evidence extraction from policy documents in Task 2.3, and D2.8 on runtime data extraction in Task 2.5. All these details contributed to D2.1 on the overall information model of the certification graph in Task 2.1.