ScholarWorks@성균관대학교

초록

In machine learning systems, detecting data drift is essential to maintain model accuracy and reliability over time. This paper presents an efficient method for data drift detection using in-stream data from pre-trained vision models, such as ResNet and Vision Transformer (ViT). Rather than processing raw input data, our approach leverages internal features extracted during inference, minimizing computational and memory overhead. We conducted experiments using six pre-trained models on the ImageNet-1k dataset, simulating data drift by introducing Gaussian noise with a variance of 0.1. The drift detection model was trained separately for each vision model and achieved high detection accuracy, ranging from 98.1% to 99.9%. Moreover, the lightweight design of the drift detection model ensured efficiency, as demonstrated by the low parameter ratio relative to the pre-trained models. Our results confirm that this approach is both effective and scalable, particularly in resource-constrained environments. © 2025 IEEE.

키워드