Abstract
Deepfake technology has enabled the widespread distribution of manipulated facial content online, raising serious societal concerns. In recent years, deepfake detection has emerged as a critical research focus. However, existing methods frequently overlook the connection between local details and overall image features, while also failing to address the problem of implicit identity leakage. Consequently, their performance is suboptimal, particularly in cross-dataset evaluations. Specifically, the proposed multi-attention deepfake detection model consists of the following three parts: (1) Texture Feature Enhancement: We employ CondenseNet to enhance texture features efficiently, preserving subtle details and ensuring feature integrity; (2) Multi-Scale Artifact Detection: We introduce an artifact detection module that identifies potentially manipulated regions, enabling localized detection and minimizing the impact of identity information. (3) Multi-Attention Mechanism: By generating multiple attention maps, our model prioritizes different regions of the input image, fusing both texture and local features to improve classification performance. Our method is evaluated on the FaceForensics++ and DFDC benchmarks for facial manipulation detection. Additionally, we assess its cross-dataset performance on Celeb-DF-v2, achieving state-of-the-art results.