Drift Type and Magnitude Detection in Image Classification Neural Networks
Keywords: Artificial Intelligence, Machine Learning, Image Processing, Classification Neural Networks, Drift Data Detection, Image Noise Level Estimation
TL;DR: This paper proposes a framework for diagnosing data drifts in the input stream of image classification neural networks due to various effects.
Abstract: A change in the input data stream of a machine-learning model is referred to as a data drift and may impact the model’s accuracy. This paper proposes a framework to detect data drifts, identify the type of drift, and estimate the drift magnitude that occur in the input data stream of image classification neural networks due to various effects. It applies to any type of drift that occurs in images due to various factors such as noise, weather, etc. A novel statistical method is proposed for drift magnitude estimation. The method relies on the change in the prediction probability distributions of the predicted classes in the classification network caused by the data drift. The drift magnitude is estimated by applying a set of thresholds to the prediction probabilities. The drift type is identified using a classification neural network. Experimental results obtained using various datasets, drift types, and neural network architectures show that the proposed framework can accurately detect data drifts, accurately identify the drift type, and estimate the drift magnitude with a very low quantization error.
Primary Area: applications to computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 13441
Loading