It was normally considered that seeing is believing until we learnt that photo editing tools can be used to alter the images we see. Technology has taken this one notch higher where facial expressions of one person can be mapped onto another in realistic videos known as deepfakes. However, each of these manipulations is not conclusive as all the image and video editing tools leave traces to be identified.
A group of researchers led by Video Computing Group of Amit Roy Chowdhury at the University of California, Riverside has created a deep neural network architecture which can detect the manipulated images at pixel level with very high accuracy. Amit Roy Chowdhury is a professor of computer science and electrical engineering at Rosemary Bourns College of Engineering. He is also a Bourns Family Faculty Fellow. The study has been published in the IEEE Digital Library.
As per artificial intelligence researchers, the deep neural network is a computer system which has been trained to perform specific tasks which, in this case, identify the altered images. The networks are organised in several connected layers.
Objects which are present in images have boundaries and whenever an object is removed or inserted to an image, its boundary will be different than the boundary which is normally present. People having good Photoshop skills will try their best to make these boundaries look natural. Examining pixel by pixel brings out the differences in the boundaries. As a result, by checking the boundaries, a computer can distinguish between a normal and an altered image.
Scientists labelled the images which were not manipulated and relevant pixels in the boundaries of the altered images in a large photo dataset. The neural network was fed the information about manipulated and the natural regions of the images. Then it was tested with a training dataset of different images and it could successfully detect the manipulated images most of the times along with the region. It provided the probability of the image being a manipulated one. Scientists are working with still images as of now, but this technique can also be used for deepfake videos.
Roy Chowdhury pointed out that a video is essentially a collection of still images, so the application for a still image will also be applied to a video. However, the challenge lies to figure out if a frame in a video is altered or not. It is a long way to go before deepfake videos are identified by automated tools.
Roy Chowdhury pointed out that in cybersecurity, the situation is similar to a cat mouse game, with better defence mechanisms the attackers also come up with better alternatives. He pointed out that a combination of a human and automated system is the right mix to perform the tasks. Neural networks can make a list of suspicious images and videos to be reviewed by people. Automation can then help in the amount of data to be sifted through to determine if an image has been altered or not. He said that this might be possible in a few years’ time with the help of the technologies.
Journal Reference: IEEE Digital Library