Published: Feb. 3, 2023

Daniel Acuña, Department of Computer Science, University of Colorado Boulder

Robust Scientific Image Tampering Detection through Noise Inconsistencies and Null Models

Scientific image tampering is a critical issue that undermines the research community's perception of trust and integrity. Current methods for identifying image tampering in natural images may not be suitable for scientific images, which have unique statistics, formats, and intentions. To address this problem, we propose a novel method for scientific image tampering detection that combines noise inconsistencies and null models of image reuse. Our method utilizes a scientific-image specific tampering detection algorithm based on noise inconsistencies to learn and generalize to different fields of science. We train and test our method on a new dataset of manipulated western blot and microscopy images, which aims to emulate problematic images in science. The test results show that our method can detect various types of image manipulation in different scenarios robustly, and it outperforms existing general-purpose image tampering detection schemes. Additionally, we estimate a null model of scientific image reuse using a high-dimensional density estimation of ORB features, providing a p-value for image reuse during research integrity investigations. Our method can provide meaningful feedback during research integrity investigations by providing a null hypothesis for scientific image reuse and thus a p-value during deliberations. Applications beyond these two types of images and future work are also discussed.