엣지 장치용 경량 LLM 한국어 스팸 탐지 추론 성능 비교 = Inference Comparison of Lightweight LLMs for Korean Spam Detection on Edge Devices

다국어 초록 (Multilingual Abstract)

This study aims to identify the optimal lightweight language model (LLM) for real-time Korean spam SMS detection on the Jetson Orin Nano edge platform. The methodology utilizes a legal-category dataset of 148,937 samples and a specialized four-stage Korean preprocessing pipeline—comprising morphological analysis, surface normalization, dictionary-based conversion, and tokenization— established in a prior study. Three lightweight models, Gemma3-1B, TinyLlama-1.1B, and DeepSeek-1.3B, were comparatively evaluated using Macro-F1 and per-category Recall across four legal categories defined under Article 50 of the Act on Promotion of Information and Communications Network Utilization and Information Protection.Experimental results demonstrate that Gemma3-1B achieved a Macro-F1 of 0.927 and an 'Illegal Activity' category Recall of 0.941, effectively minimizing False Negatives for high-risk violations while maintaining a Perplexity increase within 4.3% after INT4 quantization. The performance gap between Gemma3-1B and TinyLlama-1.1B was statistically significant (paired t-test:    , Hedges'    ). These findings confirm that Gemma3-1B is the most suitable model for Korean spam detection systems, as it balances classification accuracy, legal-category reliability, and edge deployability.