Blair, a content creator, presents her content on the digital gaming video-streaming platform Twitch. She is reasonably popular with over 800,000 followers. At the end of last January, Blair was surprised by the spread of a pornographic video clip in which she appeared. Blair was pretty sure it wasn’t her in the video, though the face of the performer was almost certainly hers. She wasn’t the only one in that shocking and frightening situation. Several female content creators on the same platform faced a similar ordeal on the same day. Behind the spread of these porn videos was a Twitch creator, Brandon Ewing, better known as Atreoc. In Ewing’s live broadcast the day before, some of his followers had noticed that in the background of the image on a computer screen appeared a porn site. They could tell that the clip on the displayed page showed one of the female content creators on the platform. Tracking the site revealed that Ewing had purchased services from this site. These services consisted of creating video clips that use images of specific women in addition to pornographic videos, with the faces of actresses in the original clips replaced with the faces of women from their photos.
Cutting and pasting faces between digital photos is an old practice. For decades, since the first graphics software appeared, this software has been used to create fake images by isolating a person’s face from an image and then replacing another person’s face in a different image with the face of the first person. With the advent of video editing software, this practice has moved to video clips. But in all cases, producing elaborate fakes required special skills in the use of specialized photo and video editing software. No matter how elaborate the work was, many people could distinguish fake photos from authentic ones. But this goes back to a time before AI-based tools entered the world of photo and video editing. What Blair and her colleagues were exposed to was a relatively recent form of the use of these technologies to create fake videos. That form that began to emerge in 2017 is called “deep fake”, a suggestive name befitting the dazzling and terrifying possibilities it reveals day after day at the same time.
What is a deepfake?
The term “deepfake” is a combination of two words: “Fake” which reflects that this type of content is not real, and “Deep” which reflects its penetration down to the smallest detail, which makes it more capable of deceiving the recipient than any previous technology. In fact, the (Deep) part was taken from the term “Deep Learning”, which is one of the research and applied fields branching out from Machine Learning, which is specifically concerned with working to enable machines (usually computers) to use huge packages of data to learn to perform certain tasks as efficiently as possible. The field of machine learning is independent of the field of artificial intelligence, but it intersects with it specifically when the tasks that it seeks to enable machines to perform simulate the performance of humans so that it is difficult to distinguish between the results of what they achieve and what humans can achieve when performing the same task.
Relentless research in the field of artificial intelligence has made successive important breakthroughs in the last decade. Those breakthroughs have begun to revive research into simulating the Neural Networks that make up the brains of humans and animals with numerical mathematical models. The idea behind this simulation is as old as the start of research in the field of artificial intelligence itself and goes back to the pioneers of this field in the fifties of the last century. The first attempts all failed and reached a dead end, to the extent that researchers in the field of artificial intelligence agreed to abandon these attempts and consider the idea itself as a sort of science fiction that is unattainable in reality. That all changed in 2012 when a team of Canadian researchers led by Geoffrey E. Hinton succeeded in producing a digital model simulating neural networks that succeeded for the first time in carrying out the content classification task of the world’s largest digital image library at the time, outperforming, by a wide margin, existing technologies that do not rely on artificial neural networks. Hinton and colleagues’ work paved the way for the rapid development of deep learning.
The next critical step in advancing the development of deep learning technologies was taken by a team of researchers led by Ian Goodfellow with their invention of the Generative Adversarial Network (GAN) in 2014. Goodfellow and his co-authors’ invention was responsible for a paradigm shift in the field of deep learning and was the gateway through which the technologies of this field lead to the production of successive generations of artificial neural network models, with the capabilities of each generation doubling compared to the previous one. Through this very gate passed the technologies of the so-called deepfake. In order to understand how these technologies differ from those that preceded them, a simple definition should be provided of the way in which competitive generative networks operate.
An adversarial generative model is composed of two artificial neural networks. One of them plays the role of the generator, which tries to learn to generate an image that simulates as accurately as possible an original element that is presented to it through a huge amount of data, which are digital images in which this element appears repeatedly in different states and shapes. In turn, the competing network plays the role of the discriminator, which learns to distinguish between the original component and the simulation produced by the generator. This competition between the two networks continues in the form of a zero-sum game, that is, the success of one of its parties means the failure of the other until the discriminator fails to differentiate between the original element and its simulation. At this point the generator would have attained the most accurate simulation, that is, as per the size of data and the size of the network measured by the number of artificial neurons constituting it, as well as the number of variables through which it operates.
The breakthrough achieved by generative adversarial networks has specifically opened the door to generic generative models that are not developed from the outset to work on a specific task or set of tasks, but rather depend on specializing in achieving a specific goal on the data that is provided to it to learn to generate data in a similar pattern, and not necessarily identical to any of provided element. What this means is that these models or algorithms developed through them can not only distinguish an item (image, written or heard word, etc.) that they have previously encountered, but can also distinguish an item that is similar to it according to specific criteria. If these networks are trained with a large number of car images, they can recognize a car in an image they have not encountered before, and if they are asked to produce an image of a car according to certain specifications, then they will produce an image of this car according to the required specifications with an accuracy that depends on the size of the data used to train it, and the number of variables and specifications that can be used.
The above is precisely what makes deepfakes different from any previously produced fake content. When a deepfake application that has been trained on a large number of images of women is asked to generate a new image of a woman based on an original image of her with the addition of some modifications, it relies on all the data it learned previously to make these modifications as accurately as possible. This means, for example, that it is able to strip a woman of her clothes based on pictures of her in full clothes, relying upon generating the parts that the clothes hide on the totality of the data that it learned. It can also superimpose the face of a woman on the body of another with the necessary adjustments according to the required scenario for facial expressions, direction, etc., which makes it ideal for generating not only still images, but also video clips with increasing accuracy from one generation to the next for these applications.
Sensity, a company that specializes in distinguishing original content from that generated by deep fake technologies, published a report titled “The State of Deepfakes: Landscape, Threats, and Impact”. The report attempts to provide an approximate picture of the rate of development of technology based on generative adversarial networks by monitoring the number of research related to this technology each year from 2014 until the year in which the report was issued, which is 2019. According to this monitoring, the growth of the field began with three research in 2014, this number developed to 9 research in 2015, then 74 in 2016, and jumped to 469 and 932, in the following two years, to reach 1207 research issued in 2019. In parallel with this development, deepfake applications began to spread in turn since they first appeared at the end of 2017. The same report provides a rough estimation of the increasing spread of these applications through the growth of the popularity of one of its models, a project called Faceswap, whose name literally means “exchange of faces”. The project is open source software and, like many other early deepfake applications, is based on the TensorFlow software development platform developed by the Google Brain team and released as open source and free software in 2015.
How is deepfake used to target women?
Deepfakes are not, in and of themselves, misusing AI technologies for criminal or immoral purposes. There are, in fact, dozens of actual and potential applications of deepfake software for legitimate and beneficial purposes. However, the vast majority of deepfakes fall under the category of immoral acts, which are, or should be, criminalized. According to reports issued by specialized research centers, 95% of monitored deepfakes were non-consensual pornographic content, meaning that it was produced without the consent of those whose images were exploited in it. On the other hand, there is nothing about the technologies used to produce false content that makes using them to target women easier than using them to target men. However, the vast majority of those whose images were exploited in deepfakes are women. More than that, the work of the agencies that worked and is working on producing deep fakes for the benefit of others, whether for a fee or for free, is almost limited to what targets women alone.
Deepfake has evolved with the evolution of the technologies on which it is based. With the increasing capacity of these technologies, as well as with the emergence of application programming interfaces (APIs) for them that make their use more accessible, the scope of their targeting of women has expanded. At the beginning of the emergence of deepfake content, it was almost limited to targeting celebrities, especially American movie stars. This is because the use of deepfake tools at the beginning required more specialized technical knowledge, and then the commercial exploitation of these tools required a high percentage of demand against supply, which is only available when a large number of those wishing to obtain it accept this product, which is not achieved through the limited social circles of ordinary women, but is only achieved in the case of celebrities who are the focus of interest of hundreds of thousands willing to obtain pornographic materials in which they appear. On the other hand, producing satisfactory results required providing the algorithms used with a huge amount of data (images and videos), which is not sufficiently available except for celebrities.
In two articles she published on the Motherboard website, in December 2017 and January 2018, Samantha Cole monitored the spread of fake pornographic videos in which famous movie stars and singers appeared. The makers of these clips superimposed the celebrities’ faces on the bodies of actresses in the original clips. These stars include Scarlett Johansson, Taylor Swift, and Gal Gadot. Describing Gadot’s video in her first article, Cole asserted that the clip “wouldn’t fool anyone if they looked closely at it.” She pointed out a number of visual defects that could reveal the falsity of the clip. But at the same time, she admitted that for the majority of those who might watch it, it would seem very real and believable. These videos were produced and posted by someone who created a community on Reddit called Deepfakes, in the first use of the term deepfake, which has become the name used to refer to this type of content in both mass media and academic research.
Cole obtained comments from the creator of the Reddit community, who declined to reveal his real identity and used the name Deepfakes instead. In his statements, Deepfakes explained that he had used open source software libraries such as Tensor-Flow and Keras, in addition to the Google image search engine, and other sources and videos on YouTube. And by training the software on each of the photos and videos of the targeted actress or singer on the one hand, and on pornographic videos on the other for a sufficient time, it becomes possible, he says, to direct the software to carry out a specific task, which in this case is to create a video with the specifications of pornographic clips, with the replacement of the face of the performer in it with the face of the target star.
In her second article, Cole monitored a significant expansion of deepfake content production with the development of a software called FakeApp. The purpose of the software, as its developer told Cole, is to make deepfake technology available to the average user without the need to write complex code to use the available software libraries. This rapid development was an indication first of the extent of demand for this type of software, and secondly, it was a warning of how easy it is to spread highly efficient tools to produce false content that is easy to be deceived by it. Although since the beginning of the emergence and spread of deepfake content, it was clear that its producers tended to target women almost exclusively, the media coverage and the attention of political circles in the United States first and then in the rest of the Western countries were directed to the political and security threats of the spread of this content, and in particular, the effect it could have if used at election times to deceive voters and direct their vote.
Targeting women has been the primary defining feature of deepfake content from the start. In a third article, Cole documented the widespread use of deepfake software by regular and non-professional users to produce fake pornographic videos in which they replaced the faces of female performers with the faces of their friends and colleagues at university and at work. Sensity’s aforementioned report confirms that all pornographic content that could be monitored was directed exclusively against women. While the majority of non-pornographic fake content was directed against men, with a huge difference in the ratio of the first type of content to the total compared to the second type.
Although there are no recent statistics similar to those presented by the Sensity report, it is still possible to estimate the extent of the expansion and inflation of the phenomenon of producing pornographic content directed against women using two basic indicators. The first is the development of related technologies at an accelerated rate in recent years, and the emergence of many specialized websites on the Internet specialized in providing the average user with pornographic content that does not need more than one photo of the target victim, based on which the website produces fake content for her within seconds. The second indicator is more obvious because it includes a specific number. In 2019, Sensity reported a total of nearly 14,000 porn videos distributed among 9 deepfake content websites, and on the top 10 most visited porn websites. In the current year, a report by Sensity also monitored that a bot on the Telegram chat platform had alone produced more than 100,000 fake porn videos. This is a very significant indicator, and it confirms that the current volume of fake pornographic content is certainly very large, which means exposing the lives of at least hundreds of thousands of women to very serious threats.
Currently, deep learning applications at the forefront are the likes of ChatGPT, which is based on Generative Pre-trained Transformer technology, developed by OpenAI; BERT models by Microsoft, LaMDA by Google, etc. Among the specialized applications for image synthesis are Midjourney by the company of the same name, and DALL-E from Open-AI. This software is a different generation of deep learning software, and it is more advanced than its predecessors. Most of them rely on a database that has the entire content of the entire Internet, and it works to convert text descriptions into images (and more recently, video clips as well). Hence, its user only needs to type a few words to produce extremely realistic images that are very difficult to distinguish from the original photos.
The deepfake impact on its victims
The emergence and spread of pornographic images or videos of any woman, whether these images and clips are real or fake, is a matter that causes long-term effects, and may accompany the victim throughout her life. These effects vary in type, between material and moral, and in terms of severity, according to many social and economic conditions of the victims, especially with different societies and cultures. But in general, it can be said that there is no society, no matter how liberal or tolerant it is, in which the reputation and social consideration of the victim would not be affected by this type of targeting. These effects are always reflected on the way the narrow social circles and society in general treat the victim.
The economic material effects may lead to the loss of the victim’s job or the cessation of her self-employment, which means the loss of her source of income. Whether the administration of the victim’s workplace or the business partners are convinced of the authenticity or falsity of the pornographic content, wouldn’t change their belief that the spread of this content permanently stigmatizes the victim in a way that may have negative effects on the business. This mainly relates to the prevailing social dynamics that turn the victim of any transgression of a sexual nature into a sexual object, meaning that she, as a person, is almost completely absorbed in the incident of the spread of pornographic content attributed to her. This becomes what determines how others deal with her to the extent that many can no longer treat her as a colleague, work partner, classmate, or otherwise. This creates a suffocating environment for the victim in her place of work or study, even if she did not lose either of them directly, in many cases she is forced to take the decision to give up work or study to avoid this suffocating environment that persecutes her with the crime of which she was the victim.
The effects on the security and safety of the victim are more serious, and it is also an extension of her definition through the crime that was committed against her as a sexual object. The chances of the victim being subjected to verbal and physical sexual harassment and attempts at sexual assault that may amount to rape, double. The published materials are used to stalk her, as a kind of calibration, humiliation, and breaking of her will in order to push her to submit to immoral initiatives and pursuits. Particularly in many Third World countries, family and neighborhood circles pose more acute sources of danger to the safety and possibly the life of the victim. It is difficult for many who lack sufficient knowledge of the advanced technology of artificial intelligence to believe that the pornographic content attributed to the victim is false and has nothing to do with it, and therefore the victim may be exposed to abuses that may lead to her killing in light of the traditions of protecting reputation and honor prevalent in many conservative and traditional societies. Which is increasing in rural and remote areas.
Last, but by no means least, the psychological effects on the victim represent the huge, submerged part of the iceberg. These effects go deeper than the victim herself can realize unless she resorts to professional help. On a conscious level, the victims likened the shock of the emergence of pornographic content that included their images to physical sexual abuse. This analogy is not an exaggeration. Every violation of the body’s integrity, even if it is known to be in reality someone else’s body, thus it remains a symbolic expression of the body itself, creates a crack in the psychological integrity of the victim. The direct material violation is equal to the indirect violation. The effects and psychological damage are exacerbated by the stance of the social circles surrounding the victim, which rarely provide real support in the case if this type of assault, but rather the opposite prevails in an extension of a prevailing culture, perhaps in the societies of the whole world, that does not vary among them in terms of type, but only in terms of degree. This culture adopts different ways of blaming the victim. In all cases, it is believed that it is in the nature of males in society to seize any opportunity to obtain any kind of sexual benefit from women in their surroundings or whom they usually know because of their fame. This belief involves the perception that a man cannot be blamed for his instinctive behavior, while his victim can be blamed for allowing the opportunity to exploit her with any behavior, even if it is at the core of her personal freedom and does not cause any harm to others, and even if it is as simple as publishing personal pictures in which the victim appears in full clothes.
Ways to combat deepfake
Much of the research work directed at combating the effects of deepfakes has focused on automated recognition of fake content using algorithms that recognize whether a particular photo or video has been modified. Some researchers believe that Blockchain technology has the potential to restore trust in the digital environment and help combat the deepfake threat. There is a conviction of the need to integrate the efforts of the various parties to support digital governance to protect victims who are subject to violations of their right to privacy and physical and psychological integrity.
Big tech companies like Google, Twitter and Facebook have taken some action to counter deepfakes, with the number of fake images doubling this year on the internet. Since 2019, Google has made freely available data that can help identify fake content research. At the same time, several specialized companies offer tools to detect fake content that can be accessed on the Internet. But the fact is that none of these tools can detect all forms of deepfakes. Some of these tools use the same technologies used by deepfake software. Despite the theoretical capabilities of this approach, it is flawed in that it needs prior training that is only available from a dataset that is identical or at least similar to that on which the software producing the fake content trained, and the collection of all these datasets is impossible. On the other hand, other tools try to use innovative approaches that focus on known vulnerabilities of deepfake software in producing some details realistically, but again, these approaches do not succeed in dealing with content that does not include those details clearly or does not include them at all.
On the legal side, there are no laws specifically designed to combat deepfakes, other than laws passed so far by three US states. In October 2019, the governor of the US state of California ratified an anti-deepfake law, in anticipation of the next elections year, specifically in an attempt to prevent false content from being used to discredit candidates in these elections. Brandie Nonnecke, in an article dating back to November 2019, presents four drawbacks to the law that, in her opinion, make it incapable of achieving its goal. These drawbacks are timing, determining criminal responsibility, the burden of providing evidence of guilt, and inadequate treatment. The law limits its criminalization of publishing false content to the 60-day period preceding the holding of any elections, it exempts sites that allow the publication of false content from liability, places the burden of providing evidence of guilt on the victims, and in the end, it does not provide adequate treatment or acceptable redress for the harm caused to the victim.
Besides Nonnecke remarks, this law, like others, reflects politicians’ apprehension of the political and security implications, and they completely lose sight of the responsibility to protect the party most affected by the vast majority of deepfake content, which is women who are targeted all the time with this content. This absence of a clear political will, even in one of the largest democracies in the world, to confront a serious threat to the security and safety of women, reflects, in fact, the absence of societal will. The world’s societies are still not at all ready to deal with the repercussions of the rapid development of artificial intelligence technologies, which necessarily affects the most vulnerable groups in society, especially women. The real ways of confrontation should go to the social roots of the phenomenon. Obsession with women’s bodies, while men express it clearly through practices such as the production of fake porn content, is an obsession rooted in mainstream cultures and is not directed by the desire to satisfy a sexual instinct as prevails in the public imagination, but rather by a belief that women’s bodies are still ingrained as objects of ownership. The real source of all forms of violence against women, which take old, usual, or new forms that exploit the latest technological and scientific achievements of human beings, is the desire to own women’s bodies and the stubborn refusal of their exclusive right to own their own bodies. All images of infringement on women’s bodies are images of stripping them of control over their bodies and of owning them, even for a fleeting moment, in a false way. Confronting society with its shameful sexism is the only way to confront all forms of violence against women at their roots and sources. In the end, it should be remembered that any technology reflects the prevailing social and cultural biases in the society that produces it, and is not in itself a source of violence, but rather a tool for it.
This paper attempted to provide a definition of what deepfake is and the technological basis that allowed its development to its current forms and which still presents great opportunities for the further development of this phenomenon. The paper sheds light on the almost exclusive targeting of deepfake technologies in its pornographic aspect of women rather than men. It also exposed the development of the size and nature of targeting in recent years. The paper tried to clarify the extremely dangerous effects of the false content on the life and safety of its victims, with a focus on the psychological effects. The paper also dealt with technological and legal efforts to try to combat the illegal uses of these technologies and explained their extreme shortcomings in light of the absence of a political will, which is a reflection of the absence of a social will. The emphasis here remains that a phenomenon of this growing magnitude has been growing for years, at a time when it has witnessed extensive discussions in the media, academia, and parliamentary legislative assemblies in the West, but there is hardly any talk about it in the Arab world, although the penetration of the use of the Internet in the region leaves no room to suspect the existence of this phenomenon locally.