Localization of Synthetic Manipulations in Western Blot Images

TL;DR

The paper tackles local synthetic manipulations in Western blot images by adopting a patch-wise detection approach. Each $P \times P$ patch is evaluated by a CNN detector to produce a patch score, which is assembled into a tampering heatmap $ ext{H}$ that highlights synthetic regions; the method remains robust to unknown generators. Two public datasets are introduced: automatically manipulated and realistically tampered Western blots, with comparisons to the TruFor baseline showing competitive localization performance and very low false alarms. Generalization is further explored on the M3Dsynth dataset, revealing some drop in performance due to semantic differences and unseen generators but indicating promising cross-domain robustness. Overall, the work provides a practical, scalable means to localize synthetic content in scientific images and contributes valuable datasets for continued forensic evaluation.

Abstract

Recent breakthroughs in deep learning and generative systems have significantly fostered the creation of synthetic media, as well as the local alteration of real content via the insertion of highly realistic synthetic manipulations. Local image manipulation, in particular, poses serious challenges to the integrity of digital content and societal trust. This problem is not only confined to multimedia data, but also extends to biological images included in scientific publications, like images depicting Western blots. In this work, we address the task of localizing synthetic manipulations in Western blot images. To discriminate between pristine and synthetic pixels of an analyzed image, we propose a synthetic detector that operates on small patches extracted from the image. We aggregate patch contributions to estimate a tampering heatmap, highlighting synthetic pixels out of pristine ones. Our methodology proves effective when tested over two manipulated Western blot image datasets, one altered automatically and the other manually by exploiting advanced AI-based image manipulation tools that are unknown at our training stage. We also explore the robustness of our method over an external dataset of other scientific images depicting different semantics, manipulated through unseen generation techniques.

Figures (8)

• Figure 1: Representation of the localization problem to identify regions of manipulation. The tampering mask $\mathbf{M}$ is shown as a reference.
• Figure 2: Pipeline of proposed methodology. We start from an image $\mathbf{I}$, whose patches are extracted sequentially and given as input to the synthetic image detector. The detector outputs probability scores that are aggregated to obtain the estimated heatmap $\mathbf{H}$, localizing the region of tampering. The tampering mask $\mathbf{M}$ is shown as a reference.
• Figure 3: Pristine examples from the Western blots dataset released in Mandelli2022.
• Figure 4: Synthetic examples from the Western blots dataset released in Mandelli2022.
• Figure 5: Localization results on automatically tampered Western blot images, for a patch size of $32 \times 32$ pixels and stride $4 \times 4$.