Affd-Net: Attention-Based Multi-Stream Feature Fusion Network to Robust Cross-Dataset Overview of Deepfake Detection

Detection of deepfakes has become a more difficult task due to the Escalating Sophistication of Reproductive Reproductions, particularly D-F architectures, such as the existing methods, which have problems with cross-dataset generalization because they rely on single-stream deep features and naive concatenation approaches. In this Paper, we present AFFD-Net (Attention-Guided Feature Fusion Detection Network), a new lightweight multi-stream model to achieve powerful deepfake detection. AFFD-Net concurrently derives complementary information in three streams: (1) deep semantic features using MobileNetV2, (2) texture features by HOG and LBP, and (3) frequency-domain artifacts with Discrete Cosine Transform (DCT). These streams are with dynamism combined with a CAG, which dynamically weights the contribution of each stream to each input sample. Extensive experiments on the 140K R-F-F datasets show that AFFD-Net achieves 99.70% validation accuracy. More to the point, it shows great zero-shot cross-dataset generalization, as it achieves 92.07% accuracy on the CIFAKE dataset, which consists of images generated by Stable Diffusion. These findings identify the usefulness of multi-domain feature fusion and attention-based dynamic weighting to forgery-type-agnostic deepfake detection.