A Deepfake Detection Method Based on Transformer Transformer Spatiotemporal Modeling

This study proposes a novel spatiotemporal feature modeling model based on the Transformer architecture–STFormer–for deepfake detection tasks in videos. The model design features a four-stage pyramid structure, each incorporating a Spatiotemporal Embedding Layer (STEL) and multiple Transformer Encoder blocks. STEL extracts spatial features using multi-scale convolutional kernels and employs a Gate Recurrent Unit (GRU) to model inter-frame dynamic changes, ultimately fusing spatiotemporal features to form embedded representations. Furthermore, the model progressively reduces the number of embedding tokens at each stage while increasing the feature dimensions, thus efficiently extracting spatiotemporal features while maintaining information integrity and effectively managing computational budgets. Experimental validation on the FaceForensics++ and Celeb-DF datasets demonstrates that the STFormer model performs exceptionally well in detecting various forgery methods, substantiating its outstanding performance and broad applicability in deepfake detection.