Learning Local Reconstruction Errors for Face Forgery Detection

Although several deepfake detection technologies have achieved great detection accuracy inside the data domain in recent years, there are still limitations in cross-domain generalization. This is due to the model’s ease of fitting the data sample distribution in the training data domain and its tendency to detect a specific forgery trace in order to reach a judgment rather than catching generalized forgery traces. In this paper, we propose to learn Local Reconstruction Errors for face forgery detection. The local anomaly traces of the fake face are often mapped using the original real face as a reference; however, the original real face of the fake face cannot be acquired in the real scenario. Therefore, this solution designs a local reconstruction autoencoder trained with real samples. By masking key areas of the face, the original real face can be reconstructed. Because the autoencoder only learns how to restore the essential parts of the real face using local patches of real samples, it cannot recover the forging traces or target face information in the fake face. Therefore, the reconstructed image forms a reconstructed difference with the original image. This solution aids the model in detecting local differences in fake faces by producing feature-level local difference attention mappings in the network’s middle layer. A series of experiments demonstrate that this solution has good detection and generalization performance.

@article{Wu2024,
  title     = {Learning Local Reconstruction Errors for Face Forgery Detection},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {15},
  number    = {11},
  year      = {2024},
  publisher = {The Science and Information Organization},
  author    = {Haoyu Wu and Lingyun Leng and Peipeng Yu},
  doi       = {10.14569/IJACSA.2024.01511120},
  url       = {https://doi.org/10.14569/IJACSA.2024.01511120}
}