Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Which Model Generated This Image? A Model-Agnostic Approach for Origin Attribution

Fengyuan Liu, Haochen Luo, Yiming Li, Philip Torr, Jindong Gu

TL;DR

This work addresses origin attribution of AI-generated images when only a few examples from a source model are available and the model itself is not accessible. It introduces OCC-CLIP, a CLIP-based framework that treats origin attribution as a few-shot one-class classification problem, using learnable prompts and adversarial data augmentation to separate target model-derived images from open-domain negatives, with fixed CLIP encoders. The approach extends to multi-source attribution via One-vs-Rest and demonstrates strong performance across eight generative models and real-world DALL·E-3 API tests, outperforming baselines such as zero-shot CLIP and CoOp and showing robustness to shot count and image processing. Practically, OCC-CLIP provides a model-agnostic, scalable method for IP protection and accountability in AI-generated content, with potential applicability to other domains beyond image synthesis.

Abstract

Recent progress in visual generative models enables the generation of high-quality images. To prevent the misuse of generated images, it is important to identify the origin model that generates them. In this work, we study the origin attribution of generated images in a practical setting where only a few images generated by a source model are available and the source model cannot be accessed. The goal is to check if a given image is generated by the source model. We first formulate this problem as a few-shot one-class classification task. To solve the task, we propose OCC-CLIP, a CLIP-based framework for few-shot one-class classification, enabling the identification of an image's source model, even among multiple candidates. Extensive experiments corresponding to various generative models verify the effectiveness of our OCC-CLIP framework. Furthermore, an experiment based on the recently released DALL-E 3 API verifies the real-world applicability of our solution.

Which Model Generated This Image? A Model-Agnostic Approach for Origin Attribution

TL;DR

This work addresses origin attribution of AI-generated images when only a few examples from a source model are available and the model itself is not accessible. It introduces OCC-CLIP, a CLIP-based framework that treats origin attribution as a few-shot one-class classification problem, using learnable prompts and adversarial data augmentation to separate target model-derived images from open-domain negatives, with fixed CLIP encoders. The approach extends to multi-source attribution via One-vs-Rest and demonstrates strong performance across eight generative models and real-world DALL·E-3 API tests, outperforming baselines such as zero-shot CLIP and CoOp and showing robustness to shot count and image processing. Practically, OCC-CLIP provides a model-agnostic, scalable method for IP protection and accountability in AI-generated content, with potential applicability to other domains beyond image synthesis.

Abstract

Recent progress in visual generative models enables the generation of high-quality images. To prevent the misuse of generated images, it is important to identify the origin model that generates them. In this work, we study the origin attribution of generated images in a practical setting where only a few images generated by a source model are available and the source model cannot be accessed. The goal is to check if a given image is generated by the source model. We first formulate this problem as a few-shot one-class classification task. To solve the task, we propose OCC-CLIP, a CLIP-based framework for few-shot one-class classification, enabling the identification of an image's source model, even among multiple candidates. Extensive experiments corresponding to various generative models verify the effectiveness of our OCC-CLIP framework. Furthermore, an experiment based on the recently released DALL-E 3 API verifies the real-world applicability of our solution.
Paper Structure (29 sections, 4 equations, 7 figures, 32 tables, 1 algorithm)

This paper contains 29 sections, 4 equations, 7 figures, 32 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Approach
  4. Background of CLIP-based Classification
  5. CLIP-based Few-shot One-Class Classification
  6. CLIP-based Few-shot Multi-Class Classification
  7. Experiments
  8. Experimental Setting
  9. Comparison with Baselines
  10. Ablation Study
  11. Comparison to Standard Data Augmentation
  12. Understanding Adversarial Data Augmentation
  13. Source Model Attribution with Commercial Generation API
  14. Model Attribution with Multiple Source Models
  15. Conclusions
  16. ...and 14 more sections

Figures (7)

  • Figure 1: A simple demonstration of origin attribution in a practical, open-world setting. Inspectors receive a few samples from DALL·E-3. Users then submit an image, which could either be a real photograph or generated by DALL·E-3 or other models. If it's determined that the image and the provided samples were generated by the same model, we can then identify DALL·E-3 as the origin model of the query image.
  • Figure 2: Overview of OCC-CLIP. The input text is represented by learnable context vectors, followed by two discrete classes: the target class corresponds to an image set queried from a generative model, and the non-target class corresponds to randomly sourced open-domain images. These classes can be labeled as contrasting pairs, such as non-target vs. target or negative vs. positive. The parameters for the text and image encoders, derived from the CLIP model, are fixed. Adversarial Data Augmentation (ADA) calculates the gradient of each pixel across non-target images. In the training phase, these gradients $\delta^v$ are applied to the non-target images.
  • Figure 3: Evaluation of OCC-CLIP on Different Numbers of Shots. This figure shows the average origin attribution performance of CoOp and OCC-CLIP on 7 different testing tasks with a variable number of shots: 10, 20, 50, 100, and 200. The target dataset is from SD. The non-target datasets are from COCO, CC12M, Flickr, or ImageNet.
  • Figure 4: Visualize the 2-dimensional mapping of image features with t-SNE van2008visualizing. The target images are from SD, and the non-target images are from COCO. Figure (a) shows the distribution of non-target images and target images without using ADA. Figure (b) shows the distribution of augmented non-target images and target images after using the ADA technique.
  • Figure 5: This figure presents a demo of synthesized images produced by six distinct models, each predicated upon four specific captions. Caption 1: Birds perch on a bunch of twigs in the winter. Caption 2: A coffee table sits in the middle of a living room. Caption 3: Three teddy bears, each a different color, snuggling together. Caption 4: The dinner plate has asparagus, carrots and some kind of meat. Caption 5: A large body of water sitting below a mountain range.
  • ...and 2 more figures