MDL-Net: Multi-Task Learning Network for Face Forgery Detection and Localization Using Dual-Stream Feature Extraction and Reconstruction

Abstract

Advancement in face forgery technologies has led to the generation of increasingly realistic images, making them difficult for the human eye to distinguish. These technologies pose risks in areas such as face recognition and identity verification, including fraud and social engineering attacks. While face forgery detection algorithms achieve high accuracy within a single dataset, their generalization performance in cross-dataset scenarios remains insufficient. To address this issue, this study introduces the Multi-task Face Forgery Detection and Localization Network (MDL-Net), which improves cross-dataset detection performance through dual-stream feature extraction and reconstructed feature enhancement. First, the face forgery reconstruction module restores and reconstructs images, amplifying subtle features in the tampered regions. Second, leveraging pixel-level residual information, the reconstruction attention module enhances feature representation by comparing input images with their reconstructed counterparts. Finally, the localization module precisely identifies forged regions through pixel-level segmentation. This approach encourages the model to extract forgery-specific features and learn more generalized and robust representations, achieving high detection and localization accuracy. Experiments conducted on the FaceForensics++ and WildDeepfake datasets demonstrate the superior performance and robustness of the proposed framework. Comparative experiments on other benchmark datasets validate the enhanced generalization ability of the method while maintaining high detection accuracy. The code is available at https://github.com/jinxinhuo/MDL-Net