IPD-Net: Detecting AI-Generated Images via Inter-Patch Dependencies

With the rapid development of generative models, the fidelity of AI-generated images has almost reached a level that is difficult for humans to distinguish true from fake. The rapid development of this technology may lead to the widespread dissemination of fake content. Therefore, developing effective AI-generated image detectors has become very important. However, current detectors still have limitations in their ability to generalize detection tasks across different generative models. In this paper, we propose an efficient and simple neural network framework based on inter-patch dependencies, called IPD-Net, for detecting AI-generated images produced by various generative models. Previous research has shown that there are inconsistencies in the inter-pixel relations between the rich texture region and the poor texture region in AI-generated images. Based on this principle, our IPD-Net uses a self-attention calculation method to model the dependencies between all patches within an image. This enables our IPD-Net to self-learn how to extract appropriate inter-patch dependencies and classify them, further improving detection efficiency. We perform experimental evaluations on the CNNSpot-DS and GenImage datasets. Experimental results show that our IPD-Net outperforms several state-of-the-art baseline models on multiple metrics and has good generalization ability.

@article{Chen2024,
  title     = {IPD-Net: Detecting AI-Generated Images via Inter-Patch Dependencies},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {15},
  number    = {7},
  year      = {2024},
  publisher = {The Science and Information Organization},
  author    = {Jiahan Chen and Mengtin Lo and Hailiang Liao and Tianlin Huang},
  doi       = {10.14569/IJACSA.2024.01507133},
  url       = {https://doi.org/10.14569/IJACSA.2024.01507133}
}