Magnitude Estimation of Multiple Data Drifts for Image Classification Neural Network

Data drift, caused by changes in environment from training, can significantly degrade the performance of neural networks in real-world computer vision tasks. Traditional drift detection methods often focus on a single type of drift, limiting their robustness in complex scenarios. In this paper, we extend an existing method that estimates the degree of a single drift for an image classification model by targeting multi-drift problem for which a set of answer keys is prepared for combination of drift levels and the answer keys are used to predict the magnitudes of drift combination applied to the images. We evaluated our extended method using the CIFAR-10 dataset with two drifts of various levels of Gaussian noise and rain effects. Experimental results confirmed that predicting multiple drifts together effectively prevents the overestimation of drift magnitudes. Additionally, the proposed drift detection model further improves accuracy of drift magnitude estimation.