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

Authors: Fred Grabovski, Lior Yasur, Guy Amit, Yisroel Mirsky

Published: 2024-07-21 13:58:43+00:00

AI Summary

This paper introduces Back-in-Time Diffusion (BTD), a novel unsupervised anomaly detector specifically designed for medical deepfakes. BTD re-purposes diffusion models to detect synthetic content by performing a single reverse diffusion step on a suspected image without adding noise, and then analyzing the residual. The method significantly outperforms existing state-of-the-art unsupervised detectors in detecting fake tumors injected into and removed from CT and MRI scans.

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.

Key findings

BTD significantly outperforms other state-of-the-art unsupervised detectors across various medical imaging modalities (CT and MRI) and deepfake attacks (injection and removal). It achieved an average AUC of 0.9 for injection deepfakes, up from 0.79 for the next best method, and an average AUC of 0.96 for removal deepfakes, up from 0.91. The method demonstrates robustness across different imaging scanners and deepfake technologies.

Approach

The proposed BTD method is an unsupervised anomaly detector that leverages Denoising Diffusion Probabilistic Models (DDPMs). Unlike previous diffusion-based anomaly detection methods, BTD directly denoises the target image for a single backward diffusion step without first adding any noise. The anomaly score is then calculated as the squared magnitude of the residual difference between the original and the single-step denoised image, which highlights anomalous forensic content.

Datasets

Duke Breast Cancer MRI dataset, LIDC-IDRI dataset, and six newly created medical deepfake evaluation datasets (CTGAN-CT-Inject, CTGAN-CT-Remove, SD-CT-Inject, SD-CT-Remove, SD-MRI-Inject, SD-MRI-Remove).

Model(s)

Denoising Diffusion Probabilistic Models (DDPMs), specifically a U-Net-based Diffusion model.

Author countries

Israel

← Previous