AdfNet: An Adaptive Deep Forgery Detection Network Based on Diverse Features

The harm caused by video tampering has been endangering people’s lives, which makes deep forgery detection technology gradually obtain widespread attention and development. However, current detection methods could not effectively capture noisy residuals due to the use of inflexible constraints. In addition, they ignore the correlation between texture and semantic features and the impact of temporal features on detection performance improvement. To solve these problems, this paper proposed an adaptive network (AdfNet) with diverse features for deep forgery detection. It helps the classifier to judge authenticity by extracting semantic features, texture features and temporal features. The paper explored the adaptive texture noise extraction (ATNEM) mechanism, and flexibly captured the noise residuals in non-fixed frequency bands through unpooled feature mapping and frequency-based channel attention mechanism. The deep semantic analysis guidance strategy (DSAGS) was designed to highlight the tampering traces through spatial attention mechanism, and guide the feature extractor to focus on the deep features of the focus region. The paper studied multi-scale temporal feature processing (MTFPM), and used temporal attention mechanism to assign weights to different video frames and capture the difference of time series in tampered videos. The experimental results show that the ACC score of the proposed network in the HQ mode of FaceForensics++ (FF++ ) dataset is 97.41%, which is significantly better than that of the existing mainstream algorithms. Moreover, while maintaining the AUC value of 99.80% on the FF++ dataset, the AUC value can reach 76.41% on Celeb-DF, reflecting strong generalization.