Attention Map Modification to Improve Deepfake Detection in Unseen Datasets

DOI

Bibliographic Information

Other Title
  • Attention Map修正による未学習のデータセットに対するDeepfake検出性能向上

Abstract

Deepfakes pose serious privacy and disinformation threats. Previous deepfake detection models often suffer from low detection accuracy on unseen datasets that are not used in training. While some studies introduce attention mechanisms to improve explainability and detection accuracy, such models tend to focus on unmanipulated areas, still causing limited accuracy on unseen datasets. This paper proposes a novel method to improve the performance of attention-based deepfake detection models on unseen datasets. Based on self-blended images, we use manipulated area maps as supervision signals of learned attention maps so as to a network is trained to focus on the manipulated areas. We evaluate the performance of the proposed method by testing it on four datasets not used for training. The results show that the proposed method outperforms baseline models by about 2% in the AUC score, achieving the best performance over the state-of-the-art methods.

Journal

Keywords

Details 詳細情報について

Report a problem

Back to top