Robust Attribution of Synthetically Generated Western Blot Images

Abstract

Retracted papers commonly include manipulations of images and figures that are unfit for publication. While some manipulations are benign, like increasing contrast or zooming in, others are designed to fool the intended audience. Recent improvements in generative computer vision pose a security risk to scientific review since generated images are often indistinguishable from authentic bioscience evidence; even to field experts. In this work, we improve upon previous attempts to detect synthetic images by attending to their differences in the frequency domain. Additionally, we solve the multiclass classification of synthetic Western blots and attribute Western blot images to their respective generative model architecture. We demonstrate that our method outperforms previous methods for synthetic Western blot detection; including efforts to classify JPEG compressed images

Motivation

Figure 1 shows that the average western blot patch from Mandelli et al. and the average fourier spectra. While there is often little difference between the average western blot patches, the average fourier spectra reveal that there are differences in high frequency pixel patterns.

Motivated by this observation, we propose a method to detect synthetic western blot images by attending to their differences in the frequency domain. We demonstrate that our method outperforms previous hand crafted and deep learning methods for synthetic western blot detection.

Method