Reduced Spatial Dependency for More General Video-level Deepfake Detection