‘Deepfakes’ refer to the practice of superimposing audio or video over an existing source
The new generation of deepfakes are more dangerous and difficult to spot
Take a look at this list: Barack Obama, Vladimir Putin, Mark Zuckerberg, Morgan Freeman, Kim Kardashian and Mamata Banerjee. These are just some of the famous names whose faces have been morphed successfully in videos and images with the help of Artificial Intelligence (AI). Not just that. While some of these faces were swapped digitally and placed on someone else’s body in explicit videos, others have been manipulated to spread misinformation.
The term “deepfakes", a portmanteau of “deep learning" and “fake", first came into existence in 2017 after an anonymous Reddit user, “Deepfakes", posted several explicit videos of celebrities on the internet. This controversial and increasingly dangerous technique of audio and video manipulation can be used to superimpose existing images or videos of a person on another source image or video. What began as a tool to make fake celebrity pornographic videos is now being deployed as a tool to spread hoaxes and generate fake news.
Till now, the threat was believed to affect celebrities and public figures whose photographs and videos are easily available in the public domain. But the debate took a turn recently with the emergence of the DeepNude app. The app, designed by anonymous developers based in Estonia, used AI to create fake nude photographs of women. With the help of neural networks, a user could virtually replace a person’s clothes with a nude body. Available for Windows and Linux, it was taken offline after a public uproar.
The app’s founders put out a statement on Twitter last week that “despite the safety measures adopted (watermarks)… the probability that people will misuse it is too high. The world is not yet ready for DeepNude," the statement added.
A key component that completes the vicious cycle of deepfakes is the AI system known as GANs, or generative adversarial networks. GANs are high-level machine-learning systems that were initially designed for “unsupervised learning"—where the AI learns on its own. But over the years GANs have also been used to create fake images and videos. The photographs generated are often realistic.
According to a report released last year by the Netherlands-based company DeepTrace, which designs technologies for detecting fake videos, GANs represent the most sophisticated recent development in how new kinds of synthetic media are created.
While the report adds that public awareness about deepfakes has grown—with Google searches for deepfakes being 1,000 times higher in 2018 than in 2017—it also adds that deepfakes are “likely to have a high profile, potentially catastrophic impact on key events or individuals in the period 2019-2020".
“We definitely have come across morphed images here in India and edited or spliced videos but no outright use of a deepfake video," says Sagar Kaul, founder, MetaFact, a Delhi-based tech-media startup that tackles fake news using fact-checking and content validation tools. MetaFact offers a fact-checking tool (in the form of an AI-based chatbot) that detects, monitors and counters fake news or misinformation in real time.
Spotting a fake
In a research paper (“Mal-Uses Of AI-generated Synthetic Media And Deepfakes: Pragmatic Solutions Discovery Convening") published in July 2018, WITNESS, an international non-profit organization that encourages people to use video and technology to protect and defend human rights, listed out possible solutions to detect deepfakes. These ranged from basic steps such as checking the source of the information to spotting the absence of blinking in deepfakes (computer-generated faces do not blink as often). Another basic flaw in deepfakes is the length of the clip: Deepfake video clips are usually short—not longer than a few seconds.
“When the first generation of deepfakes came into the public space, the golden rule for detecting them was the eye blinks. Since the technology used in creating the videos was still at an early stage, training tools to detect them seemed easy. But with the second generation of deepfakes, the videos created are more sophisticated and harder to detect—both with the naked eye or applications, because of the more advanced GANs used in creating such content," Kaul explains.
People who create deepfake content of world leaders and celebrities have access to a huge database of their images and videos. Not only can the creator train a machine-learning model using this database of audio and visual data, they can also train GANs to learn how to evade detection from other software or applications.
Fighting this remains a challenge. Kaul says sharing such data sets (images, videos and audio) between researchers, journalists, fact-checkers and technology companies might help create an ecosystem to build better detection tools.