Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Addressing Vulnerabilities in AI-Image Detection: Challenges and Proposed Solutions

Justin Jiang

TL;DR

This work addresses the challenge of reliably detecting AI-generated images in the face of evolving generative models. It systematically evaluates CNN-based detectors and DenseNets on CIFAKE-derived datasets produced by Stable Diffusion across versions, with variations such as Gaussian blur, LoRA, and prompt changes. The results reveal strong version-specific overfitting and vulnerability to common degradations, while DenseNets offer improved robustness and higher accuracy. The findings highlight the need for diverse, version-robust training data and architecture choices to build detectors that generalize in real-world, evolving-generation environments.

Abstract

The rise of advanced AI models like Generative Adversarial Networks (GANs) and diffusion models such as Stable Diffusion has made the creation of highly realistic images accessible, posing risks of misuse in misinformation and manipulation. This study evaluates the effectiveness of convolutional neural networks (CNNs), as well as DenseNet architectures, for detecting AI-generated images. Using variations of the CIFAKE dataset, including images generated by different versions of Stable Diffusion, we analyze the impact of updates and modifications such as Gaussian blurring, prompt text changes, and Low-Rank Adaptation (LoRA) on detection accuracy. The findings highlight vulnerabilities in current detection methods and propose strategies to enhance the robustness and reliability of AI-image detection systems.

Addressing Vulnerabilities in AI-Image Detection: Challenges and Proposed Solutions

TL;DR

This work addresses the challenge of reliably detecting AI-generated images in the face of evolving generative models. It systematically evaluates CNN-based detectors and DenseNets on CIFAKE-derived datasets produced by Stable Diffusion across versions, with variations such as Gaussian blur, LoRA, and prompt changes. The results reveal strong version-specific overfitting and vulnerability to common degradations, while DenseNets offer improved robustness and higher accuracy. The findings highlight the need for diverse, version-robust training data and architecture choices to build detectors that generalize in real-world, evolving-generation environments.

Abstract

The rise of advanced AI models like Generative Adversarial Networks (GANs) and diffusion models such as Stable Diffusion has made the creation of highly realistic images accessible, posing risks of misuse in misinformation and manipulation. This study evaluates the effectiveness of convolutional neural networks (CNNs), as well as DenseNet architectures, for detecting AI-generated images. Using variations of the CIFAKE dataset, including images generated by different versions of Stable Diffusion, we analyze the impact of updates and modifications such as Gaussian blurring, prompt text changes, and Low-Rank Adaptation (LoRA) on detection accuracy. The findings highlight vulnerabilities in current detection methods and propose strategies to enhance the robustness and reliability of AI-image detection systems.

Paper Structure

This paper contains 22 sections, 1 equation, 4 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. CIFAKE Dataset and Classifier
  4. Methods
  5. Experiments
  6. Datasets
  7. Results
  8. Replicating CIFAKE Method.
  9. Evaluation Across Stable Diffusion Versions.
  10. Evaluation on Gaussian Blur.
  11. Evaluation on Image Size Sensitivity.
  12. Evaluation on Prompt Variability.
  13. Evaluation on Low-Rank Adaptation (LoRA) Alternation.
  14. Using DenseNets to Detect Stable Diffusion-Generated Images.
  15. Discussion
  16. ...and 7 more sections

Figures (4)

  • Figure 1: CIFAKE Dataset structure.
  • Figure 2: Structure of the DenseNet model used in our experiments. The model includes an initial convolutional layer, followed by multiple dense blocks and transition layers, concluding with a final linear layer for binary classification.
  • Figure 3: Example images from the CIFAKE huggingfaceDragonintelligenceCIFAKEimagedatasetDatasets dataset.
  • Figure 4: Example image from the CIFAKE-SD2.1-LoRA dataset with its photorealism trigger word.