Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

EditTrack: Detecting and Attributing AI-assisted Image Editing

Zhengyuan Jiang, Yuyang Zhang, Moyang Guo, Neil Zhenqiang Gong

TL;DR

EditTrack tackles the problem of detecting whether a suspicious image is derived from a base image via AI editing and attributing the editing model. It introduces a re-editing strategy that uses a caption-derived proxy prompt $p_e$ to generate two re-edited images per candidate model, and then leverages six similarity metrics to form $12n$ features fed into an $n+1$-class classifier, enabling both detection and attribution without access to model parameters. Across five editing models and multiple datasets, EditTrack achieves high detection and attribution accuracy and consistently outperforms baselines, with ablations confirming the importance of both re-edited image groups and all similarity metrics. This post-hoc forensic approach provides a practical and scalable tool for tracing AI-assisted image edits and could be extended to other modalities such as text and video in future work.

Abstract

In this work, we formulate and study the problem of image-editing detection and attribution: given a base image and a suspicious image, detection seeks to determine whether the suspicious image was derived from the base image using an AI editing model, while attribution further identifies the specific editing model responsible. Existing methods for detecting and attributing AI-generated images are insufficient for this problem, as they focus on determining whether an image was AI-generated/edited rather than whether it was edited from a particular base image. To bridge this gap, we propose EditTrack, the first framework for this image-editing detection and attribution problem. Building on four key observations about the editing process, EditTrack introduces a novel re-editing strategy and leverages carefully designed similarity metrics to determine whether a suspicious image originates from a base image and, if so, by which model. We evaluate EditTrack on five state-of-the-art editing models across six datasets, demonstrating that it consistently achieves accurate detection and attribution, significantly outperforming five baselines.

EditTrack: Detecting and Attributing AI-assisted Image Editing

TL;DR

EditTrack tackles the problem of detecting whether a suspicious image is derived from a base image via AI editing and attributing the editing model. It introduces a re-editing strategy that uses a caption-derived proxy prompt to generate two re-edited images per candidate model, and then leverages six similarity metrics to form features fed into an -class classifier, enabling both detection and attribution without access to model parameters. Across five editing models and multiple datasets, EditTrack achieves high detection and attribution accuracy and consistently outperforms baselines, with ablations confirming the importance of both re-edited image groups and all similarity metrics. This post-hoc forensic approach provides a practical and scalable tool for tracing AI-assisted image edits and could be extended to other modalities such as text and video in future work.

Abstract

In this work, we formulate and study the problem of image-editing detection and attribution: given a base image and a suspicious image, detection seeks to determine whether the suspicious image was derived from the base image using an AI editing model, while attribution further identifies the specific editing model responsible. Existing methods for detecting and attributing AI-generated images are insufficient for this problem, as they focus on determining whether an image was AI-generated/edited rather than whether it was edited from a particular base image. To bridge this gap, we propose EditTrack, the first framework for this image-editing detection and attribution problem. Building on four key observations about the editing process, EditTrack introduces a novel re-editing strategy and leverages carefully designed similarity metrics to determine whether a suspicious image originates from a base image and, if so, by which model. We evaluate EditTrack on five state-of-the-art editing models across six datasets, demonstrating that it consistently achieves accurate detection and attribution, significantly outperforming five baselines.

Paper Structure

This paper contains 12 sections, 4 equations, 5 figures, 11 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Problem Formulation
  4. EditTrack
  5. Overview
  6. Extracting Features
  7. Training a Multi-class Classifier
  8. Evaluation
  9. Experimental Setup
  10. Main Results
  11. Ablation Studies
  12. Conclusion and Future Work

Figures (5)

  • Figure 1: Three similarity categories.
  • Figure 2: Validation of the four observations. $\uparrow$ / $\downarrow$ indicate that higher / lower values correspond to greater similarity.
  • Figure 3: Ablation studies.
  • Figure 4: Image pair samples from different datasets. The first row shows base images, and the second row shows their corresponding suspicious images.
  • Figure 5: Image samples generated using different editing models. The first column shows the base images. The editing prompts are: first row-"Do the image editing task; origin prompt: two elephants playfully interact while splashing through a muddy waterhole in a lush, green landscape, editing prompt: two rhinoceros playfully interact while splashing through a muddy waterhole in a lush, green landscape"; second row-"Do the image editing task; origin prompt: a space shuttle launches dramatically amidst billowing smoke and towering clouds against a clear sky, editing prompt: a helicopter rises swiftly amidst swirling dust and towering clouds against a clear sky".

Theorems & Definitions (3)

  • Definition 1: AI-assisted Image Editing
  • Definition 2: Image-Editing Detection
  • Definition 3: Image-Editing Attribution