Microsoft Debuts A Weapon To Combat deepfakes in Elections

Videos, with real images, are artificially modified by an algorithm and often manipulate the speech of a public person. Among the efforts that technology companies are making to combat misinformation at this pre-election juncture in the United States, Microsoft has taken an important step. It has presented a platform apparently capable of detecting deepfakes.

The artificial intelligence behind these manipulated videos makes it difficult to identify their falsehood. It is possible that a text read by an actor goes through some statements by Barack Obama , as happened in a video that tried to raise awareness about this technique. But it is also possible to change the face of the protagonist of a video for that of another person. Thus, Jim Carrey replaces Jack Nicholson in this deepfake of The Shining .

The Microsoft Video Authenticator tool uses its own artificial intelligence to analyze the veracity of a video. “It evaluates the existence of areas of the image in which there may be a change in textures, areas blurred by the superposition of elements, consistency of the edges, alteration of colors”, explains Alberto Pinedo, Director of Technology at Microsoft in Spain . They look for “traces of a modification on the original content that normally escape the human eye”.

In each frame, the Microsoft algorithm estimates a probability that the images are real or manipulated. The tool is also used to verify the authenticity of photographs. In the case of videos, the verification is done in real time throughout the duration of the footage.
For now, Microsoft Video Authenticator will be available through the Reality Defender 2020 (RD2020) initiative, a platform that brings together academic entities, such as the University of Berkeley, and technology companies, Google, Twiter, Microsoft itself, to combat the disinformation. In addition, the company will collaborate with the BBC in its Project Onion, aimed at combating fake news globally and with an eye toward the next US elections.

The Microsoft system has been created using two public databases: Face Forensic ++, to train the algorithm, and DeepFake Detection Challenge Dataset, to test its effectiveness. Although the company admits that its Video Authenticator is not the ultimate weapon against deepfakes . “It is a platform in continuous evolution and learning to improve its precision. We must bear in mind that we are faced with mechanisms that improve every day to create better quality fake videos ”, Pinedo emphasizes.

Easy access from the cloud, technology deep learning ( neural networks needed to create these manipulations ) allows almost anyone to perpetrate a deepfake . This type of video is already proliferating on YouTube. On many occasions they are harmless jokes, simple memes in audiovisual format, like this clip in which Jon Snow apologizes for season 8 of Game of Thrones . But they can also be part of intentional hoaxes.

The origin of deepfakes can be anyone, as Pinedo points out, although he emphasizes the incidence of social networks and online video platforms. Precisely these two spaces have been vehicles for the dissemination, sometimes viral, of hoaxes and misinformation. A problem that with the manipulation of videos is complicated.

And it is that detecting them with the naked eye is not easy. From the Massachusetts Institute of Technology (MIT) they propose to look at some details to try to identify a deepfake . It involves discovering irregularities in the skin of the face, on the cheeks and forehead, in the color and size of the lips, the haircut. Is there something weird about it? And in the blink, is it too frequent or hardly exists? There are other aspects to pay attention to : if we detect obvious failures, such as a shirt collar that is one way on the left side and another, the alarms should go off.

But recognizing these types of details requires a close examination, not feasible on a day-to-day basis. Especially at the rate that digital content is consumed today. Hence, some important Internet technologies have taken steps to combat manipulated videos in an automated way. Last year Google launched a database with 3,000 deepfakes generated with artificial intelligence, to contribute to research in this field. For its part, Facebook has stimulated a competition to detect manipulated deep videos . Precisely the database that is used as a test in this contest has been the one that Microsoft has used to test its algorithm.