By Hannah Kolev
This is the third post in a series about artificial intelligence, along with its uses and social/political implications.
An enchanting video of the Spanish artist Salvador Dalí - who passed away in 1989 - remarking on the current weather with visitors of the Dalí Museum. A video of British soccer star David Beckham speaking nine different languages - only one of which he actually speaks - petitioning world leaders to end malaria. A faked pornographic video used to discredit investigative journalist Rana Ayyub for speaking out against a sex abuse case. In each of these scenarios, deepfake videos were used to dupe their audiences. These deepfakes can be used to ignite the imagination, to inspire change, or to intimidate victims. But what are deepfakes and how are they made? What is their purpose and how are they regulated? We will address these questions in the following blog post.
Deepfakes, a combination of the terms “deep learning” and “fake”, are videos generated using artificial intelligence (AI) to produce fake images. First introduced in a 2017 Reddit post, deepfakes morph the face of one person onto another person’s body. In some cases, deepfake technology can even generate realistic faces of people who do not exist. In 2018, a striking example of a deepfake video was produced by Buzzfeed News in which former US President Barack Obama insulted then President Donald Trump. In this deepfake, a previously recorded video of President Obama was manipulated and merged with a video recording of actor Jordan Peele speaking the insult. The result was a hyper-realistic deepfake of President Obama speaking the words of Jordan Peele.
To produce this type of “face-swapping” video, deep learning algorithms and neural networks are used to superimpose the facial features of a “target” person onto the face of a “source” person. The resulting video shows the target person (i.e., President Obama) performing the actions of the source person (i.e., Jordan Peele). Specifically, two autoencoders – a type of artificial neural network used to learn image representations – are trained on input data, which typically include hundreds of images of the faces to be learned. The first autoencoder is trained to recognize the target person, while the second autoencoder is trained to recognize the source person. Importantly, each autoencoder can be decoupled into two separate neural networks: an encoder and a decoder. The encoder extracts latent facial features from the images, while the decoder then reconstructs the face. During the training process, the encoder parameters are shared between the two autoencoders, allowing the encoder to learn common features between the target and the source. Once training is complete, the deepfake is created by feeding the output of the encoder for the source into the decoder for the target, thereby producing a new autoencoder. When this new autoencoder is passed a video of the source person, it outputs a video of the target person, thus creating the face swap. Together, this creates the illusion that the target person is performing the same actions as the source person in the original video.
More recent technological advancements have employed a type of deep learning AI called a “generative adversarial network” (GAN) to produce deepfake videos. In this case, two artificial neural networks called the “generator” and the “discriminator” are trained on the same dataset of images or videos. The generator then creates new images, which are fed into the discriminator. The discriminator’s job is to determine whether these new images are real or fake. In this way, the two neural networks compete and push each other to improve. Through this iterative improvement process, GANs can produce incredibly accurate deepfakes.
Beyond producing jarring videos of former presidents, deepfakes have potential benefits for various industries including media and cinema, marketing and advertising, and healthcare. Firstly, the film industry is poised to embrace deepfake technology, given its potential applications for quickly and realistically dubbing movies into multiple languages, creating movies starring previously deceased actors, and “de-aging” actors to enable them to play younger versions of themselves. Blockbuster movies such as Captain Marvel and Rogue One have already created “de-aged” versions of actors, although these movies typically rely on time consuming and costly technologies, such as computer-generated imaging, instead of deepfakes. However, as deepfake technology continues to improve, it offers a viable option for quickly and easily producing hyper-realistic “de-aged” videos of actors.
Deepfake technology also has implications for marketing and advertising, with many companies using deepfakes as a personalization tool for brands. Given that consumers are more likely to make a purchase when provided with a personalized shopping experience, companies are using deepfakes to create highly targeted and individualized shopping environments. In the clothing industry, companies are starting to use deepfakes to allow shoppers to visualize different models, or even themselves, in different clothing options before purchasing them. For example, the startup company Tangent is using GANs, one of the technologies behind deepfakes, to modify the faces of models in an effort to produce catalog images for consumers in different countries. This process side-steps the need to hire different models or to use other image modifying software to produce catalog images that are representative of different nationalities. Even more, users of the popular deepfake app, Reface AI, were able to insert their face into a Gucci ad and virtually try on Gucci clothes, further emphasizing the utility of deepfakes in creating a personalized shopping experience. Extending beyond the clothing industry, deepfakes were also used in the 2020 Doritos Super Bowl marketing campaign featuring singer Lil Nas X. In partnership with Sway, an AI app that produces deepfake videos, Doritos created an app in which users visualize themselves dancing the same moves as Lil Nas X in his Super Bowl commercial.
Beneficial applications of deepfake technology also extend into the healthcare and medical sectors. For example, deepfake technology has been used not only to create face-swapping videos but also to create fake medical data for training AI algorithms. A 2018 study from the Mayo Clinic, in collaboration with the NVIDIA Corporation and the MGH & BWH Center for Clinical Data Science, employed GANs to create fake MRI brain scans, which were then used to train AI algorithms to detect brain tumors. They found that algorithms trained on GAN-generated images supplemented with 10% real images were just as good at identifying brain tumors as algorithms trained using solely real data. In this context, deepfakes provide a viable stand-in for real medical data and can be used to mitigate situations in which medical images are sparse, for example, in rare diseases. In the future, deepfake technology may also be applied to the digital recreation of amputated limbs, thereby helping to relieve phantom limb pains. Additional applications may include the production of deepfake videos of deceased loved ones to help family members move through their grieving process. Similarly, someone with Alzheimer’s disease may benefit from seeing a deepfake video of a younger version of themselves or someone they knew, as happy memories of the past can be soothing and may help improve their communication. Together, these examples underscore that deepfake technology, when used in benevolent and beneficial ways, can have a direct impact in society, permeating many different industries to enhance our entertainment, consumer, and healthcare experiences.
However, despite these innovative applications of deepfake technology, a significant portion of deepfakes are developed to create harm and propagate misinformation. Firstly, the vast majority of deepfake videos are pornographic in nature, often superimposing the face of a celebrity or layperson into a pornographic video. A 2019 report from Deeptrace, a cyber-security company aimed at protecting others from the damaging effects of AI-generated media, found that 96% of all deepfake videos are pornographic and nonconsensual. This type of image-based sexual abuse, also referred to as “fake porn”, is frequently produced and distributed without consent from those featured and largely targets women. I spoke with Sophie Maddocks, a PhD student in Penn’s Annenberg School of Communication, whose research focuses on cyber-sexual violence and image-based abuse. She explained that image-based sexual abuse is “so severe because it never goes away, there’s no end point.” Continuing on, Ms. Maddocks noted that the “representation of you online is real, and if it’s abused, it causes you physical, psychological, and social harm.”
Rana Ayyub, an investigative journalist, was one such victim of image-based sexual abuse. In 2018, Ms. Ayyub spoke out against India’s apparent protection of child sex abusers. In retaliation, a nonconsensual deepfake pornographic video of her was rapidly and widely disseminated online. This deepfake video was distributed in a deliberate maneuver to discredit, threaten, and harm Ms. Ayyub for her divisive opinions. Ms. Ayyub continues to suffer lasting effects of this image-based sexual abuse, as she explains in her 2018 HuffPost article: “From the day the video was published, I have not been the same person. I used to be very opinionated, now I’m much more cautious about what I post online.” Similarly, victims of deepfakes and “shallowfakes”, which are produced by manipulating already existing videos with simple editing tools, often report symptoms of mental and physical illness. Victims may also experience financial repercussions, such as unemployment. In this way, the misuse of deepfakes to create nonconsensual pornographic videos has severe, real, and lasting impacts on its victims.
In addition to their use to create image-based sexual abuse, deepfakes may also be employed for the production and dissemination of political misinformation. Politically motivated shallowfakes and deepfakes have already been widely distributed online, as former President Trump retweeted an unflattering deepfake of presidential candidate Joe Biden in the run up to the 2020 US presidential election. In 2019, a “shallowfake” of US Speaker of the House of Representatives Nancy Pelosi was created by slowing down her speech, making it seem as if she was impaired or intoxicated. And most recently, a deepfake of Ukrainian President Volodymyr Zelenskyy circulated online in which he purportedly commanded his troops to lay down their arms and surrender to Russian aggression. This video raises concerns that Russia may employ deepfakes as a method of information warfare, thereby creating panic and confusion. Although these political deepfakes are cause for concern, it is important to note that they remain far outnumbered by nonconsensual pornographic deepfakes and that their overall impact during the US 2020 presidential election was lower than original predictions. Nonetheless, the potential for misuse raises the concern that if left unchecked, political deepfakes may be used to sow confusion, create uncertainty, and lower trust in news sources.
Given the propensity for people to misuse deepfakes, it is critical for regulations and laws to be developed and enforced to thwart their abuse. Social media companies have mainly set their own regulations regarding deepfake distribution on their platforms. In 2018, companies such as Twitter and Pornhub quickly removed nonconsensual deepfake pornography, with Pornhub’s vice president condemning these deepfakes as a form of sexual abuse. Furthermore, amid growing concern about misinformation surrounding the 2020 US presidential election, Facebook banned deepfake videos that are likely to mislead voters. However, this policy does not cover shallowfakes, which are often just as misleading, and allows videos including deepfakes to remain on the site if they are deemed “newsworthy”. Twitter, on the other hand, prohibited the deceptive distribution of manipulated and fake media that may cause harm, which includes both AI-generated deepfakes and shallowfakes. On both Facebook and Twitter, posts that contain manipulated media are labeled as such, and users are warned before sharing altered or fabricated posts. This policy provides users with greater context about the videos and posts to which they are exposed - an important step in combatting the spread of misinformation and misleading deepfakes online.
In recent years, legislation bills at both the state and federal levels have also addressed deepfakes. In 2019, Virginia was the first state to pass a law criminalizing the distribution of nonconsensual deepfake pornography, with penalties including up to one year in jail or a $2,500 fine. Texas followed shortly thereafter, passing a bill prohibiting the distribution of deepfakes intended to sway elections or discredit candidates for office. And in California, a series of laws passed in 2019 permits victims of nonconsensual deepfake pornography to sue for damages and allows electoral candidates to sue those who distribute malicious deepfakes without warning labels in the lead up to elections.
At the Federal level, the National Defense Authorization Act (NDAA) for the Fiscal Year 2020 was signed into law by former president Donald Trump and outlined concrete steps to combat deepfakes. As the first federal law related to deepfakes, the NDAA outlines a three-pronged approach to combatting malicious deepfakes. Firstly, the law requires the Director of National Intelligence to submit a comprehensive report to the Congressional Intelligence Committees describing the deepfake weaponizing capabilities of foreign entities. Secondly, the law requires that Congress be notified of any foreign deepfake activities aimed at influencing US elections. Thirdly, the law creates a “Deepfakes Prize” competition with the overall goal of encouraging the research and development of technologies to detect deepfakes. The initial report, as well as updates submitted annually to Congress, should comment on the deepfake technological capabilities of foreign entities such as Russia and China and should describe how these foreign entities are using deepfakes to affect US national security. Together, these state and federal activities represent the growing interest in legislating deepfake use and distribution.
With their rapid advancement and growth, deepfakes will likely continue to play a role in our everyday lives. As Ms. Maddocks noted in our interview, “[deepfake] technology isn’t inherently malevolent.” And, as we’ve discussed here, deepfake technology has the potential to contribute to new and exciting cinematic and consumer experiences. Deepfakes can personalize our shopping and enhance medical research and healthcare practices. However, deepfakes are also prone to misuse - whether through nonconsensual deepfake pornography that frequently targets women or through the spread of misinformation in an effort to sway voters or confuse the public. When asked about the direction society is headed when considering deepfakes and, specifically, their use to create image-based sexual abuse, Ms. Maddocks explains that this “is not an easy answer and [is] definitely tied to much broader societal issues.” Ms. Maddocks notes that holistic solutions to deepfake technologies are important – solutions that consider the experiences of those who are the most oppressed and who are the most vulnerable to abuse at the hands of deepfakes. In the case of image-based sexual abuse, such holistic solutions may include educating people about privacy and about healthy interpersonal relationships. Thus, as deepfake technology continues to advance, society must advance alongside it – educating the public to understand healthy uses of deepfakes, adopting laws and policies to regulate their use and distribution, and carefully considering how deepfakes can contribute to a thriving societal ecosystem.