Self-Certification of High-Risk AI Systems: The Example of AI-based Facial Emotion Recognition
Gregor Autischer, Kerstin Waxnegger, Dominik Kowald
TL;DR
The paper addresses how to demonstrate compliance for high-risk AI under the EU AI Act by applying the Fraunhofer AI Assessment Catalogue as a self-certification framework to an AI-based facial emotion recognition system. It reconstructs EmoPy into EmoTorch, conducts a complete self-certification focused on reliability and fairness, and iteratively improves the system, achieving 68.19% test accuracy and fairer performance across demographic groups. The study shows that the catalogue drives deep technical investigation, serves as a practical development framework, and yields documentation organically, but it also confirms that self-certification cannot replace harmonized European standards or formal conformity assessment. The work highlights substantial preparatory value for future regulatory compliance while clarifying the regulatory gaps that remain, and it suggests directions for future work, including applying certification processes to real-world deployments and aligning with forthcoming standards.
Abstract
The European Union's Artificial Intelligence Act establishes comprehensive requirements for high-risk AI systems, yet the harmonized standards necessary for demonstrating compliance remain not fully developed. In this paper, we investigate the practical application of the Fraunhofer AI assessment catalogue as a certification framework through a complete self-certification cycle of an AI-based facial emotion recognition system. Beginning with a baseline model that has deficiencies, including inadequate demographic representation and prediction uncertainty, we document an enhancement process guided by AI certification requirements. The enhanced system achieves higher accuracy with improved reliability metrics and comprehensive fairness across demographic groups. We focused our assessment on two of the six Fraunhofer catalogue dimensions, reliability and fairness, the enhanced system successfully satisfies the certification criteria for these examined dimensions. We find that the certification framework provides value as a proactive development tool, driving concrete technical improvements and generating documentation naturally through integration into the development process. However, fundamental gaps separate structured self-certification from legal compliance: harmonized European standards are not fully available, and AI assessment frameworks and catalogues cannot substitute for them on their own. These findings establish the Fraunhofer AI assessment catalogue as a valuable preparatory tool that complements rather than replaces formal compliance requirements at this time.