Open Access

Philipp StraubORCiD, Christian DeckerORCiD, Marco KuhrmannORCiD

• Systems development in resource-constrained, safety-critical areas is significantly impacted by slight discrepancies, such as unknown faults, resulting in system failure. This affects life, the environment, or finances. The rise of Artificial Intelligence (AI), especially through TinyML, encourages the development of AI-based systems for dynamic fault mitigation in safety-critical domains. Architectural tactics are a crucial development practice as they address constraints and quality attributes through low-level design decisions. This paper elaborates on the state-of-the-art by identifying the quality attributes and architectural tactics used by safety-critical, resource-constrained AI-based systems. We conducted a Systematic Mapping Study using a snowballing search procedure, including 217 publications. We inductively identified 150 architectural tactics across 19 categories and provided a detailed overview. Our results show that most publications providing tactics are solution proposals, while few report long-term evaluations. This reveals an immature field. When using ISO/IEC 25059 for quality attribute classification, we found that it lacks twelve relevant attributes. Our research contributes to the development of safe and resource-friendly AI-based systems.