Researchers have developed a deep neural network architecture that can identify manipulated images at the pixel level with high precision by studying the boundaries of objects in the image.