Exposing the Fake: Effective Diffusion-Generated Images Detection

TL;DR

This work tackles the security challenge of identifying diffusion-generated images by introducing Stepwise Error for Diffusion-generated Image Detection (SeDID). SeDID exploits the deterministic reverse and denoising properties of diffusion models by analyzing the stepwise error $E_{t,\delta}$ at a chosen timestep $T_{\textit{SE}}$ with stepsize $\delta$, using two branches: SeDID_Stat and SeDID_NNs. Across CIFAR10, TinyImageNet, and CelebA, SeDID demonstrates superior detection performance compared to a diffusion-based baseline, with the neural-network variant often delivering the best results in terms of AUC and ACC. The approach advances AI security by offering a robust detector that leverages intermediate diffusion dynamics and distributional differences between real and synthesized images, and it outlines concrete directions for extending the method to broader diffusion-model families and automated parameter selection.

Abstract

Image synthesis has seen significant advancements with the advent of diffusion-based generative models like Denoising Diffusion Probabilistic Models (DDPM) and text-to-image diffusion models. Despite their efficacy, there is a dearth of research dedicated to detecting diffusion-generated images, which could pose potential security and privacy risks. This paper addresses this gap by proposing a novel detection method called Stepwise Error for Diffusion-generated Image Detection (SeDID). Comprising statistical-based $\text{SeDID}_{\text{Stat}}$ and neural network-based $\text{SeDID}_{\text{NNs}}$, SeDID exploits the unique attributes of diffusion models, namely deterministic reverse and deterministic denoising computation errors. Our evaluations demonstrate SeDID's superior performance over existing methods when applied to diffusion models. Thus, our work makes a pivotal contribution to distinguishing diffusion model-generated images, marking a significant step in the domain of artificial intelligence security.

Figures (7)

• Figure 1: The pipeline of our SeDID method. Given mixed image data, it is processed through the SeDID method to compute the noise profile, a characterization of the noise patterns inherent to diffusion-generated images. Then, SeDID provides two branches: the Statistical-Based Synthetic Image Detection $\text{SeDID}_{\text{Stat}}$ and the Neural Network-Based Synthetic Image Detection $\text{SeDID}_{\text{NNs}}$. The $\text{SeDID}_{\text{Stat}}$ branch involves statistical analysis, error calculation, and model evaluation. The $\text{SeDID}_{\text{NNs}}$ branch employs a ResNet-18 model, which computes prediction errors, and updates weights via backpropagation. Both branches calculate the Area Under the Receiver Operating Characteristic Curve (AUC), the Accuracy (ACC), and the True Positive Rate at a given False Positive Rate (TPR@FPR), and classify images with real or generated output.
• Figure 2: Effect of different Stepsize on CelebA, TINY-IMAGENET, and CIFAR10 datasets. The optimal Stepsize $\delta$ is highlighted with a triangle in each series.
• Figure 3: Visual representation of the intermediate results of the SeDID method using the STAT strategy on the CIFAR10, Tiny-ImageNet, and CelebA datasets across various Stepwise Error Calculation Time Steps, $T_{\textit{SE}}$, within the range of [0, 1000] with the stepsize $\delta$ = 165. This figure illustrates the diffusion progression under the SeDID method across different Stepwise Error Calculation Time Steps, $T_{\textit{SE}}$, at a fixed stepsize $\delta$.
• Figure 4: Performance of the SeDID method at various timesteps. This graph shows how SeDID's effectiveness varies with different time steps $T_{\textit{SE}}\in [0, 1000]$. Key observations include a smaller improvement rate in TPR@FPR(0.1%) for CIFAR10 and Tiny-ImageNet as $T_{\textit{SE}}$ increases, and the highest performance for CelebA at small $T_{\textit{SE}}$.
• Figure 5: Samples from DDPM on CIFAR-10 at step 800,000.

Theorems & Definitions (4)

• Definition 3.1: Deterministic denoising function $\psi_{\theta}$
• Definition 3.2: Deterministic reverse function $\phi_{\theta}$
• Definition 3.3: $t\text{,}\delta$-error
• Definition 3.4: Stepwise Error Calculation Time Step, $T_{\textit{SE}}$ and stepsize, $\delta$