Multisemantic Path Neural Network for Deepfake Detection

Figure 6

Network details of MSPNN. All receptive fields in micro semantic pathway are . Superposition of two convolution kernels replaces receptive fields in the mesoscopic semantic path, and skip connection is used in the second block. Two convolution kernels replace receptive fields in the macroscopic semantic path. In the third and fourth blocks, skip connections reduce loss of information. Finally, the output of each path is aligned with other semantic feature maps through downsampling.