Online Detection of AI-Generated Images

View on arXiv ← Back to list

Authors: David C. Epstein, Ishan Jain, Oliver Wang, Richard Zhang

Published: 2023-10-23 17:53:14+00:00

AI Summary

This paper investigates the online detection of AI-generated images by training a classifier on a chronologically ordered sequence of generative models. It extends this approach to pixel-level prediction for detecting inpainted regions, demonstrating the effectiveness of CutMix augmentation even without pixel-wise training data.

Abstract

With advancements in AI-generated images coming on a continuous basis, it is increasingly difficult to distinguish traditionally-sourced images (e.g., photos, artwork) from AI-generated ones. Previous detection methods study the generalization from a single generator to another in isolation. However, in reality, new generators are released on a streaming basis. We study generalization in this setting, training on N models and testing on the next (N+k), following the historical release dates of well-known generation methods. Furthermore, images increasingly consist of both real and generated components, for example through image inpainting. Thus, we extend this approach to pixel prediction, demonstrating strong performance using automatically-generated inpainted data. In addition, for settings where commercial models are not publicly available for automatic data generation, we evaluate if pixel detectors can be trained solely on whole synthetic images.

Key findings

The online classifier generalizes to unseen models, with performance increasing as the training set grows. Pixel-level detection is achievable, and CutMix augmentation improves performance significantly even without ground truth masks. However, generalization significantly degrades when architectural changes between models are substantial.

Approach

The authors train a binary classifier (ResNet-50 for whole image detection and FCN with ResNet-50 for pixel-level detection) in an online setting, sequentially adding new generative models to the training data based on their release dates. They evaluate generalization to unseen models and explore the use of CutMix augmentation for pixel-level detection.

Datasets

A dataset of 570,221 images from 14 generative models (including DDPM, DDIM, GLIDE, DALL-E 2, Stable Diffusion versions, Midjourney versions, Adobe Firefly) released between June 2020 and March 2023, along with real images from LAION-400M and synthetically generated inpainted images.

Model(s)

ResNet-50 (for whole image detection), FCN with ResNet-50 (for pixel-level detection)

Author countries

USA

← Previous