Machine Learning Model Deployment in Mobile-based Systems: State of the Art and Trends - Data extraction

Mobile technology has significantly advanced over the past two decades, transitioning from simple telecommunications devices to sophisticated smartphones, tablets, portable computers, and integrated Internet of Things (IoT) systems. In parallel, Artificial Intelligence (AI) and Machine Learning (ML) have evolved from niche research topics to ubiquitous technologies in modern society. Recent advances in mobile processing capabilities have opened new possibilities for deploying large, complex ML models from cloud servers directly to mobile devices, reducing computational costs and improving privacy and response times. Hence, understanding ML model deployment strategies for resource-constrained mobile environments has become essential, and practitioners have had difficulty deciding which strategies to adopt. This study presents the state-of-the-art and future trends for the deployment of ML models in mobile-based systems. We scrutinized the scientific literature, selected 30 primary studies, and examined them across three aspects:  deployment approaches, characteristics of mobile devices considered during deployment, and technologies supporting the deployment. Our findings reveal different deployment approaches, from simple ones (e.g., the embedded approach that has been most investigated) to those with potential for future adoption (i.e., real-time streaming and push messaging approaches still underexplored, as well as other promising new approaches). We also identified critical mobile device characteristics---processing units (CPU, GPU, and NPU), power consumption, latency, bandwidth, and memory\/storage---that should be systematically considered during design and deployment. In addition there are many supporting technologies and tools (mostly open source and community-driven), including ML model conversion tools, cloud infrastructure services, specialized network protocols, and security mechanisms, revealing significant fragmentation and lack of standardization.  Results of this study have important implications for researchers and practitioners interested in the quality of ML-enacted mobile-based systems, in terms of cost-effectiveness, sustainability, and performance.