<p>Malicious use of face forgery technology will lead to serious negative effects, so there is an urgent need for an effective face forgery detection approach. Most existing approaches directly take the cropped face regions from the whole image as input and leverage the extracted visual features to predict real and fake through a binary classifier. However, advanced forgery methods will leave no obvious visual forgery artifacts, especially under real-world high compression conditions. To address such limitations, a Dual-branch Pre-activation Bottleneck Transformer (DPBoT) is proposed for face forgery detection in this paper. Instead of directly extracting appearance features, this method designs a dual-branch model to capture both visual forgery artifacts and local noise feature inconsistencies. Furthermore, to reduce information loss and ensure smoother information propagation within the model, attention mechanisms are introduced into face forgery detection, and a novel PBoT backbone architecture is designed. Finally, features from both branches are fused by bilinear pooling. Extensive experimental results on four well-known datasets show that the proposed method achieves promising results, which outperform the state-of-the-art face forgery detection methods.</p>