Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AEDR: Training-Free AI-Generated Image Attribution via Autoencoder Double-Reconstruction

Chao Wang, Zijin Yang, Yaofei Wang, Weiming Zhang, Kejiang Chen

TL;DR

AEDR tackles image origin attribution for state-of-the-art generative models by proposing a training-free, passive method based on a double autoencoder reconstruction. It uses the ratio of two successive reconstruction losses, $t = \frac{L_1}{L_2}$, calibrated by an image homogeneity metric to form $t'$, and applies kernel density estimation to determine a threshold for belonging vs non-belonging. The approach significantly improves attribution accuracy (averaging around $95.1\%$ vs baselines) and achieves about a $100\times$ speedup by avoiding gradient-based optimization, with strong generalization across multiple autoencoder architectures and eight diffusion models. This yields scalable, practical attribution suitable for real-world deployment without modifying generation pipelines.

Abstract

The rapid advancement of image-generation technologies has made it possible for anyone to create photorealistic images using generative models, raising significant security concerns. To mitigate malicious use, tracing the origin of such images is essential. Reconstruction-based attribution methods offer a promising solution, but they often suffer from reduced accuracy and high computational costs when applied to state-of-the-art (SOTA) models. To address these challenges, we propose AEDR (AutoEncoder Double-Reconstruction), a novel training-free attribution method designed for generative models with continuous autoencoders. Unlike existing reconstruction-based approaches that rely on the value of a single reconstruction loss, AEDR performs two consecutive reconstructions using the model's autoencoder, and adopts the ratio of these two reconstruction losses as the attribution signal. This signal is further calibrated using the image homogeneity metric to improve accuracy, which inherently cancels out absolute biases caused by image complexity, with autoencoder-based reconstruction ensuring superior computational efficiency. Experiments on eight top latent diffusion models show that AEDR achieves 25.5% higher attribution accuracy than existing reconstruction-based methods, while requiring only 1% of the computational time.

AEDR: Training-Free AI-Generated Image Attribution via Autoencoder Double-Reconstruction

TL;DR

AEDR tackles image origin attribution for state-of-the-art generative models by proposing a training-free, passive method based on a double autoencoder reconstruction. It uses the ratio of two successive reconstruction losses, , calibrated by an image homogeneity metric to form , and applies kernel density estimation to determine a threshold for belonging vs non-belonging. The approach significantly improves attribution accuracy (averaging around vs baselines) and achieves about a speedup by avoiding gradient-based optimization, with strong generalization across multiple autoencoder architectures and eight diffusion models. This yields scalable, practical attribution suitable for real-world deployment without modifying generation pipelines.

Abstract

The rapid advancement of image-generation technologies has made it possible for anyone to create photorealistic images using generative models, raising significant security concerns. To mitigate malicious use, tracing the origin of such images is essential. Reconstruction-based attribution methods offer a promising solution, but they often suffer from reduced accuracy and high computational costs when applied to state-of-the-art (SOTA) models. To address these challenges, we propose AEDR (AutoEncoder Double-Reconstruction), a novel training-free attribution method designed for generative models with continuous autoencoders. Unlike existing reconstruction-based approaches that rely on the value of a single reconstruction loss, AEDR performs two consecutive reconstructions using the model's autoencoder, and adopts the ratio of these two reconstruction losses as the attribution signal. This signal is further calibrated using the image homogeneity metric to improve accuracy, which inherently cancels out absolute biases caused by image complexity, with autoencoder-based reconstruction ensuring superior computational efficiency. Experiments on eight top latent diffusion models show that AEDR achieves 25.5% higher attribution accuracy than existing reconstruction-based methods, while requiring only 1% of the computational time.

Paper Structure

This paper contains 17 sections, 6 equations, 3 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Image Generative Models.
  4. Detection of AI-Generated Images.
  5. Origin Attribution of Generated Images.
  6. Method
  7. Problem Statement and Threat Models
  8. Double-Reconstruction via Autoencoder
  9. Calibration Mechanism
  10. Threshold Determination
  11. Experiments and Results
  12. Experimental Setup
  13. Effectiveness
  14. Efficiency
  15. Generalization
  16. ...and 2 more sections

Figures (3)

  • Figure 1: Gradient-based reconstruction methods exhibit different loss distributions. These differences lead to attribution failures on the latest model, such as FLUX.
  • Figure 2: Autoencoder-based double-reconstruction loss variation and loss ratio.
  • Figure 3: The framework of AEDR. Our method consists of three key modules: double-reconstruction based on autoencoder (Section \ref{['method-1']}), calibration mechanism (Section \ref{['method-2']}), and threshold determination via kernel density estimation (Section \ref{['method-3']}). AEDR employs the ratio of double-reconstruction losses, calibrated by an image homogeneity metric, as the attribution signal. The decision threshold is determined via kernel density estimation and applied for origin attribution.