모바일 NPU 가속을 위한 AI 워터마킹 모델의 구조적 경량화 및 최적화 기법 연구

초록

This study proposes a method to optimize the performance of a deep learning-based watermarking model in a mobile Neural Processing Unit (NPU) environment. Executing the Lightweight-Mark model on the Snapdragon 8 Gen 3 NPU revealed a significant bottleneck in which the ConvTranspose2d operation accounted for 99.5% of the total inference time. To address this issue, ConvTranspose was decomposed into nearest neighbor upsampling and standard convolution, and knowledge distillation was applied to compensate for accuracy loss. An ablation study across 129 experimental configurations showed that the proposed method achieved a 98.3× speedup and a 0.67% reduction in average device temperature compared with the original Lightweight-Mark model, demonstrating the feasibility of mobile real-time watermarking.

키워드