Back-in-Time Diffusion: Unsupervised Detection of Medical Deepfakes

TL;DR

This work proposes a novel anomaly detector for medical imagery based on diffusion models and shows how a similar process can be used to detect synthetic content by making a model reverse the diffusion on a suspected image.

Abstract

Recent progress in generative models has made it easier for a wide audience to edit and create image content, raising concerns about the proliferation of deepfakes, especially in healthcare. Despite the availability of numerous techniques for detecting manipulated images captured by conventional cameras, their applicability to medical images is limited. This limitation stems from the distinctive forensic characteristics of medical images, a result of their imaging process. In this work we propose a novel anomaly detector for medical imagery based on diffusion models. Normally, diffusion models are used to generate images. However, we show how a similar process can be used to detect synthetic content by making a model reverse the diffusion on a suspected image. We evaluate our method on the task of detecting fake tumors injected and removed from CT and MRI scans. Our method significantly outperforms other state of the art unsupervised detectors with an increased AUC of 0.9 from 0.79 for injection and of 0.96 from 0.91 for removal on average. We also explore our hypothesis using AI explainability tools and publish our code and new medical deepfake datasets to encourage further research into this domain.

Figures (8)

• Figure 1: Overview of the Back-in-Time Diffusion framework. The model is a U-net that predicts the noise added in the last step ($F(x_t) \Rightarrow \epsilon_t$). The model is trained by progressively adding noise to the images $(x_0)$ to generate noisy versions $(x_t)$ and then learning to predict the noise added at each step. Detection is performed on image $x_0$ by measuring the error after $d$ steps.
• Figure 2: The number of samples from each imaging device, per dataset.
• Figure 3: Examples of manipulated images from our detection datasets. Top: original images. Bottom: images after being edited by generative models
• Figure 4: ROC Curves of BTD and the baselines organized by dataset (attack scenario).
• Figure 5: BTD Explainabiliy Experiment: BTD residuals and their SHAP values compared to residuals created with 20 F&B and their SHAP values