TY  - JOUR
T1  - Performance Analysis of Deep Learning Models for On-Device AI
AU  - Park, Jinho 
AU  - Hong, Hyuck Ki 
JO  - The Journal of Korean Institute of Communications and Information Sciences
PY  - 2025
DA  - 2025/1/1
DO  - 10.7840/kics.2025.50.12.1974
KW  - On-device AI
KW  - Deep learning
KW  - Disease diagnosis
KW  - Lung respiratory sound
AB  - Due to the increased public interest in respiratory diseases following the outbreak of COVID-19, various artificial intelligence (AI)-based disease detection studies have been actively conducted. AI-based disease detection classification by analyzing lung sounds measured through stethoscopes. Conventional AI-based detection schemes typically rely on resource-rich servers to achieve high accuracy and fast inference times.
Utilizing servers requires transmitting information such as lung sounds over a network, which raises concerns about personal data privacy. To address this issue, on-device AI―where the AI model runs locally on the device―has been gaining attention. On-device AI collects and processes data internally, thereby minimizing privacy concerns. Although various deep learning models can be deployed for on-device AI, performance degradation due to limited computing resources necessitates careful model selection. This study analyzes and evaluates the disease classification and detection performance of models executed on both server and on-device environments. Experimental results show that deep learning models have lower performance when operated on-device compared to when operated on a server.