Are “Deepfake” videos a threat to democracy and global security?
by
Scott Jenkin
Introduction:
The Collins English Dictionary added the word ‘Deepfakes’ in their 2019 edition. This word originates from the username of an anonymous person/s posting social media content. The user/s had their name enter the English language and headlines around the world because of the compromising videos of celebrities they share, or so it seemed. The videos were in fact, computer generated fakes.
In the age of meme and misinformation, this paper aims to explore what role Deepfake videos could have in shaping democracy and global security. Research for this paper begins by examining previous media deceptions, then exploring the path and process to technologies behind making a Deepfake video, while providing insight to the digital environments in which they are shared, before looking into key areas Deepfakes could disrupt democracies and threaten peace around the globe.
A brief account of media deceptions:
In modern film and television, it is unlikely for audience members to try to dodge something on screen moving towards the camera; yet a short film of a train pulling into a station had viewers clambering from their seats. The film was not in 3D, nor any trickery involved, for most of the audience, it was simply their first experience of moving image (Loiperdinger, M., & Elzer, B., 2004). The Lumière brothers did not intend for this reaction, but their then advanced technology provoked an instinctive response. In 1895 audiences were accustomed to stage shows with live performers and physical props, with no knowledge or understanding of this new technology, the impending train caused them to discard logic and react as if they were in real danger.
Almost a century on, with audiences well acquainted with cinema and live television broadcasts, the BBC televised a deception with deadly effect. In 1992, the long-standing broadcaster held an estimated viewership of 15 million during their Saturday night programming. In October of that year, they aired a special one-off drama to coincide with Halloween, Ghostwatch (1992). The 90-minute show followed a similar format to a regular, factual show, Crimewatch, a live broadcast featuring a “phone-in” for viewers to help solve crimes; this time however, investigating paranormal activity in a London home. Trusted TV personality Michael Parkinson CBE was cast to host the show from the studio, while a film crew along with Sarah Greene played out scripted events on location. By the climax of the show, the poltergeist had taken control of the camera, leaving the studio presenter seemingly perplexed. Despite the show being billed as a ‘hoax’, 18-year-old viewer, Martin Denham, believed the on-screen events; and worse, believed they were now happening in his family home. Five days after the show aired, Martin took his own life, leaving a note for his mother which read: ‘if there is a ghost I will be one and I will always be with you as one’. (Woods, 2017).
Machine Learning, AI and Deepfakes:
Artificial intelligence, also know as AI, is a term used when computing power is used as digital systems capable of completing tasks which would normally require some form of human intelligence. This could range from simple to complex operations, such as scanning email for spam, to a self-driving car scanning the road ahead to plot its journey. Machine learning is the development of these AI systems which can improve the performance of the set task through gained experience over time.
In the 1960s researchers in the field of AI predicted that a computer would beat a chess champion within a decade (Cipra, 1996). Almost three decades later, their prediction became reality with IMB’s supercomputer, Deep Blue, beating the then world champion, Garry Kasparov, twice in a six-game match in 1997. This mild success was due to Deep Blue’s processors, searching for and analysing 100,000,000 moves per second, while strategising up to 20 moves in advance. By 2003 computing had progressed, and the rematch was on, this time Kasparov set out to beat Deep Fritz and Deep Junior. These systems didn’t rely on just brute force by rapidly crunching millions of calculations, they could now learn from experiences, allowing them to account for the strategy of their human opponent. Kasparov spared defeat, holding each machine to a draw. The first victory came the follow year during the Man vs Machine World Team Championships, with both computers dispatching their human opponents with ease (Anderson, 2017).
Movie visual effects have benefited from the mathematical capabilities of AI. For the 1985 movie, Young Sherlock Holmes, the visual effects team did not have software for simulating movement of their CGI character, instead the team made the manual calculations for the computer. This painstaking method took six months to complete a shot that appeared on-screen for 10 seconds (Corridor Crew, 2019, 04:30). That same year Robert Abel was tasked with creating a 30 second CGI advert to air during the Super Bowl, Brilliance (1985). Developing a new software, the team behind the project pioneered ‘brute force animations’. The process began by painting dots on a live model to act out movements for a camera. Using the dots as a reference point, the computer ran tens of thousands of calculations for each frame, mimicking the movements and creating a vector map for the graphics to be applied. With this new AI driven approach, the project took just eight weeks to complete (MrTvRetro, 2012).
Hardware advances have meant that processing power can now be harnessed within a laptop. This has been exemplified by an amateur filmmaking crew in Nigeria using rudimentary equipment and free software to create CGI scenes unimaginable three decades ago (DeRuvo, 2019). With less limitations in the West, people are able to shoot, edit and publish entire films using just a smartphone (Lincoln, 2018). App developers have harnessed the processing power of these devices with ‘filters’, allowing users to augment selfies with novel effects, made possible through AI and machine learning. Fed and trained mass amount of data to detect human faces, the software builds a vector map and applies graphics that can be change at a swipe; all taking place in real time (Vox, 2016).
In 2014 an American AI researcher, Ian Goodfellow, had a breakthrough when trying to solve the problem of AI systems not being able to generate new, life like image by itself. He began by working on artificial neural networks, algorithms inspired by biological neural networks that function within the human brain. These deep-learning systems could be trained on data, like images, and learn from them, enabling security systems like Apple’s Face ID to recognise a user after a change in appearance. However, asking the system to create new images lead to poor results, with missing facial features or blurry images produced. By combining two neural networks in competition with each other, Goodfellow created GAN (generative adversarial network). Both networks are trained on the same data, in this case, images of human faces. One network (generator) would then be tasked with generating new data and feeding it back into the loop, while the second (discriminator) would try to identify the fake. With each cycle both networks learn and improve. Leaving the systems working longer would better the end result, in this case, producing an indistinguishable AI generated image of a human face (Giles, 2018)
GAN technology was largely limited to researchers until 2017, when a Reddit user by the name of ‘Deepfakes’ began uploading manipulated pornographic videos. Deepfakes was using open-source machine learning software from Google, TensorFlow, to create GANs which were trained to replace faces of adult film stars with female celebrities. The videos, dubbed as ‘deep fakes’, drew interest from media outlets, in response Reddit removed the content and the offender from their platform. Deep Fakes went on to release free software, FakeApp, providing the masses the ability to create fake content at will (Schwarts, 2018).
Practical Applications:
Practical applications of AI are beginning to be intertwined in our daily lives. Systems navigate us through congested traffic while we drive, or remove the clutter when we want to stream a film or listen to music. Intelligent Virtual Assistant (IVA), such as the Amazon Echo, can work in collaboration with ‘smart tech’ around the home, controlling lights, temperature and helping keep groceries replenished. Holistic AI systems are being trialled and developed for healthcare professionals. A report from the Nuffield Council on Bioethics (2018) illustrates how, in preparation, data to train AI systems has already been collected and stored. In practice the systems created have been proven capable of detecting eye diseases, pneumonia and some forms of cancer. At the time of the reports release, systems were being developed to analyse speech patterns for predicting psychotic episodes, and AI controlled robotic tools for performing surgeries.
The plethora of uses for AI and Machine Learning will continue to grow due to its versatility, trained to analyse data, recognise patterns, perform tasks and solve problems. GANs however are fundamentally limited; they analyse to synthesise, are still in relative infancy, and most of its uses are hypothetical. However, as outlined previously, video manipulation is a real application in the present. Despite Deepfakes exploitative and demeaning examples, many who have downloaded the FakeApp software have used it to recreate iconic moments from cinema; frequently with Nicolas Cage now in the starring role (Haysom, 2018).
FakeApp is proving that it can create convincing counterfeit visuals, but a persons’ likeness extends beyond appearance. Another company, Lyrebird, have used TensorFlow GANs to create a feature for their software, Descript, a system for imitating a persons’ voice. As a show of things to come, one of the founders demonstrated the beta software by sampling a few minutes of audio from a journalist to train the system. Once complete, the computer was able to deceive, via telephone, the journalists mother by mimicking his voice (Wired, 2019). Aside from impersonations, the software has a positives use, finding or recreating a voice for people who have lost the ability to speak (Maras & Alexandrou, 2019). The founders of Lyrebird are aware of the potential misuse of their software, and in an effort to combat this, they have implemented safeguards that will only allow individuals to recreate their own voice. In a statement they recognise that similar products will soon be available from other developers and that ‘… there’s no reason to assume they will have the same constraints we’ve added.’ (Descript, 2019).
Social Media and Misuse:
As of the third quarter of 2019, Facebook had 2.45 billion monthly active users (Facebook, 2019). The free-to-use platform has enabled founder, Mark Zuckerberg, to amass an estimated $68.2 billion since its inception 16 years ago (Lodenback, T, Kneven, L,, 2019).
Facebook make their fortune through data, which recently surpassed oil as the world’s most valuable resource. The data they harvest is not just the information users list in their biography, it is also data such as, cookies from other websites they have visited or real world locations their GPS enabled smartphone has frequented. Most activities now create a digital trace, building a data profile of a user. The monetising of this data is prized because of Facebook’s advanced AI algorithms which can analyse and predict when a user is in a particular mood and/or considering a purchase, and pair them with the highest bidder for their attention, providing bespoke targeted advertising (The Economist, 2017).
The AI driven algorithm does more than connecting adverts to consumers, it ensures each user an ad hoc view of the platform; photos, videos, news articles and everything else shared through the platform is analysed before being presented to an audience. In theory this saves the user from having to scroll through uninteresting content, maintaining their attention. In practice, there is a damage. For example, if a user is a supporter of the red team, the algorithm may limit favourable content for the blue team; while simultaneously connecting them to fellow supporters. Not much of a problem when red and blue represent sporting factions, but replace them with political parties and the process becomes constrained, limiting the users perceptive, showing them content which is often biased, creating what is referred to as an ‘echo chamber’ (Lanier, 2018).
AI systems were tried and tested during the 2016 Republican Party presidential primaries. It was near perfect, with the backed candid, Ted Cruz, moving his campaign from last position in the polls, to becoming the runner up. The group behind the Cruz campaign were Cambridge Analytica, whose services were enlisted by Donald Trump for his presidential election campaign. This time Cambridge Analytica would go one better, defying polls yet again to secure Trump victory and become the 45th President of the United States of America. The data-driven company also played a key role in restructuring the political landscape in Europe, by partnering with the Leave.EU campaign during the UK Brexit referendum, scoring their second major win at the ballot (The Great Hack, 2019). In an interview with the Guardian, former data scientist turned whistleblower, Christopher Wylie (2018) explains how his ex-employer spent millions of dollars harvesting data from Facebook profiles in each of the campaigns to identify undecided voters. Targeting them specifically to receive covertly placed news stories against their rivals, insidiously playing on the core concerns of those targeted. Due to the echo chambers created by the Facebook algorithm, this largely went unnoticed and therefore, mostly avoided fact checking scrutiny.
Deepfakes and Politics:
In 2018 the first known Deepfake was created and published by a political party, the Flemish Socialist party, Socialistiche Partji Anders (SP.A). The video sees President Trump making an address to the people of Belgium, giving them climate advice. Despite the video ending with the Deepfaked Trump confessing the forgery, many of those who saw the Facebook post mistook it as genuine (Burcahard, 2018).
As illustrated in the previous section, AI systems can be implemented with success in political campaigning. This section will present two scenarios in which Deepfake videos could be used with potentially devastating effects.
Scenario one:
In the eleventh hour before polling station open before an election, a Deepfake video emerges online which depicts the leader of a political party in a compromising situation and/or making offensive remarks. The video is witnessed by enough of the electorate to sway the election, influencing their vote before the video can be debunked.
The adage ‘a lie travels halfway around the world before the truth gets out of bed’ could be used succinctly to endorse the proposed scenario. With polling stations opening their doors from dawn to dusk it could be assumed that a large proportion of the electorate cast their vote before commencing their usual daily commitments; meaning that even with swift action being taken to address the fake, it might not be enough to mitigate the irreparable damage. With 49% of uk adults opting to use social media for their news consumption (Ofcom, 2019), the echo chambers syndrome created by algorithms curating what social media users see, any refutation against the Deepfake could be missed entirely. Weight is added to this scenario by US senator, Marco Rubio, when addressing a Senate Intelligence Committee hearing in May 2019, he offered the similar scenario to express concerns for Western democracies in the technological era (O’Sullivan, 2019).
As the public become more aware of GANs and how convincingly they can fabricate media content, we could see the reverse of this scenario, where a person holding a powerful position is genuinely caught in a recording, and questions the validity. The now infamous Access Hollywood audio tape where Donald Trump is heard boasting about sexually assaulting women is a prime example. At the time of its release, he immediately apologised, yet now he questions the authenticity of the tapes, claiming they have been doctored in an effort to defame him (Wall Street Journal, 2018).
Scenario two:
As explained by Chief Technologist for the University of Michigan’s Center of Social Media Responsibility, Aviv Ovadya, in an interview with BuzzFeed News (Warzel, 2018)
‘Imagine, for example, a machine-learning algorithm fed on hundreds of hours of footage of Donald Trumnp or North Korean dictator Kim Jong Un, which could then sip out a near-perfect - and virtually impossible to distinguish from reality - audio or video clip of the leader declaring nuclear or biological war. It doesn’t have to be perfect - just good enough to make the enemy think something happened that it provokes a knee-jerk and reckless response of retaliation.’
The above example of nuclear war being triggered is extremely unlikely due to safeguards in place and lines of communications open between global super-powers, however, smaller acts of war being set in motion by a misplaced hoax or a genuine attempt by a bad actor to instigate conflict should not be discounted. In an extensive report, written by a collaboration of leading experts, The Malicious Use of Artificial Intelligence: Forecasting, Prevention, and Mitigation (2018), suggests that Deepfake videos will increase due to efficiency and scalability of the technology. With just a small amount of hardware required and the software being freely available, the financial burden to start creating Deepfakes is little to none. With creative software giant, Adobe, expanding the AI tools they have to offer their clients, it could be reasonably assumed that software for creating Deepfakes is soon to follow, which will greatly increase use and misuse (Maras, A, Alexandrou, A, 2019).
Conclusion:
Beginning by looking at, how, when faced with new technologies, or if presented with authority, the receiving audience can be bamboozled by media content, expressly when the intensions are to obfuscate. The evidence throughout my research has compelled the conclusion that Deepfake videos, produced using GANs possess a real potential to create havoc for democracies and dictatorships alike. The digital environments we have created for consumption over compassion and genuine connection, allow for the often unchecked dissemination of misinformation, compounding the viability of Deepfakes as a threat. The situation is further exacerbated by the fundamentals of GAN softwares, the more we try to combat and discern their presences, the more we teach the systems to evade detection.
The Bulletin of the Atomic Scientists gave a stark warning in their 2020 Doomsday Clock Statement (2020);
‘The recent emergence of so-called ‘deepfakes’ - audio and video recordings that are essentially undetectable as false - threatens to further undermine the ability of citizens and decision makers to separate truth from fiction. The resulting falsehoods hold the potential to create economic, social and military chaos, increasing the possibility of misunderstandings to provocations that could lead to war, and fomenting public confusion that leads to inaction on serious issues facing the planet. Agreement on facts is essential to democracy and effective collective action.’.
bibliography
AdAge (1985) Brilliance. ND. Available at: https://adage.com/videos/the-canned-food-information-council-brilliance/1159 (Accessed 26 January 2020)
Africanews (2019) Nigerian teens make sci-fi films with smartphones. Available at: https://youtu.be/D25NrAQZLng (Accessed: 07 November 2019)
Anderson, M R. (2017) Twenty years on from Deep Blue vs Kasparov: how a chess match started the dig data revolution. Available at: http://theconversation.com/twenty-years-on-from-deep-blue-vs-kasparov-how-a-chess-match-started-the-big-data-revolution-76882 (Accessed: 17 January 2020)
Bulletin of the Atomic Scientists (2020) Closer than ever: It is 100 seconds to midnight. Available at: https://thebulletin.org/doomsday-clock/current-time/
Brundage, M. & Avin, S. (2018), The Malicious Use of Artificial Intelligence : Forecasting, Prevention, and Mitigation , Technical report, Future of Humanity Institute, University of Oxford; Centre for the Study of Existential Risk, University of Cambridge; OpenAI , Oxford.
Burchard, H v d. (2018) Belgain socialist party circulates ‘deep fake’ Donald Trump video. Available at: https://www.politico.eu/article/spa-donald-trump-belgium-paris-climate-agreement-belgian-socialist-party-circulates-deep-fake-trump-video/ (Accessed: 01 Sept 2019)
Cipra, B. (1996) Artificial Intelligence: Will a Computer Checkmate a Chess Champion at Last?, Science, 271(5249), pp. 599
CNBC (2019) Can Facebook And Google Detect And Stop Deepfakes?. 26 Sept 2019. Available at: https://youtu.be/4YpoYvhVmDw (Accessed: 01 Sept 2019)
CNN Business (2019) When seeing is no longer believing, Inside the Pentagon’s race against deep fake videos. Available at: https://edition.cnn.com/interactive/2019/01/business/pentagons-race-against-deepfakes/ (Accessed: 01 Sept 2019)
ColdFusion (2018) Deepfakes - Real Consequences. 28 April 2018. Available at: https://youtu.be/dMF2i3A9Lzw (Accessed: 01 Sept 2019)
Corridor (2019) Keanu Reeves Stops A ROBBERY!. 13 July 2019. Available at: https://youtu.be/3dBiNGufIJw (Accessed: 01 Sept 2019)
Corridor Crew (2019) We Made The Best Deepfake on The Internet. 20 June 2019. Available at: https://youtu.be/3vHvOyZ0GbY (Accessed: 01 Sept 2019)
Corridor Crew (2019) How We Faked Keanu Reeves Stopping a Robbery. 14 July 2019. Available at: https://youtu.be/IzEFnbZ0Zd4 (Accessed: 01 Sept 2019)
Corridor Crew (2019) VFX Artists React to Bad & Great Cgi 14. 16 November 2019. Available at: https://youtu.be/2fDuj0D6CIo (Accessed: 03 January 2020)
Corridor Crew (2019) VFX Artist React to Resurrected Actors Bad & Great CGi. 1 June 2019. Available at: https://youtu.be/2ZKPnuUFwOk?t=639 (Accessed: 27 January 2020)
DeRuvo, J. (2019) How Teenagers In Nigeria Are Making Big Moves With Smartphones. Available at: https://nofilmschool.com/nigeria-filmmakers-the-critics (Accessed: 07 November 2019)
Descript (2019) Descript Ethics Statement. Available at: https://www.descript.com/ethics (Accessed: 27 January 2020)
The Economist (2017) Regulating the internet giants: The world’s most valuable resource is no longer oil, but data. Available at: https://www.economist.com/leaders/2017/05/06/the-worlds-most-valuable-resource-is-no-longer-oil-but-data (Accessed: 12 January 2020)
Facebook (2019) Facebook Q3 2019 Results. Available at: https://s21.q4cdn.com/399680738/files/doc_financials/2019/q3/Q3-2019-Earnings-Presentation.pdf (Accessed: 21 January 2020)
Ghostwatch (1992) BBC One Television, 31 October, 21:25.
Giles, M (2018) The GANfather: The man who’s given machines the gift of imagination. Available at: https://www.technologyreview.com/s/610253/the-ganfather-the-man-whos-given-machines-the-gift-of-imagination/ (Accessed: 25 November 2019)
The Great Hack (2019) Directed by K. Amer & J. Noujaim. [Documentary film]. Los Gatos, CA: Netflix
The Guardian (2018) Cambridge Analytica whistleblower: 'We spent $1m harvesting millions of Facebook profiles’ 17 March. Available at: https://www.youtube.com/watch?v=FXdYSQ6nu-M (Accessed: 20 December 2019)
Haysom, S (2018) People are using face-swapping tech to add Nicola Cage to random movies and what is 2018. Available at: https://mashable.com/2018/01/31/nicolas-cage-face-swapping-deepfakes/?europe=true (Accessed: 28 January 2020)
Lanier, J (2018) Ten Arguments for Deleting Your Social Media Accounts Right Now. The Bodley Head: London.
Lincoln, K. (2018) How to Make a Movie on Your Smartphone. Available at: https://www.vulture.com/2018/11/how-to-make-a-movie-on-your-smartphone.html (Accessed: 25 January 2020)
Loudenback, T., Knueven, L. (2019) Mark Zuckerberg’s net worth increased By Over $1 billion after Facebook’s FTC fine - see the houses, cars, and travels where he spends is billions. Available at: https://www.businessinsider.com/facebook-mark-zuckerberg-net-worth-priscilla-chan-2017-10?r=US&IR=T (Accessed: 28 January 2020)
Loiperdinger, M., & Elzer, B. (2004). Lumiere's Arrival of the Train: Cinema's Founding Myth. The Moving Image 4(1), 89-118. doi:10.1353/mov.2004.0014.
Maras, A H, Alexandrou, A. (2019) ‘Determining authenticity of video evidence in the age of artificial intelligence and in the wake of Deepfake videos’, International Journal of Evidence & Proof, 23(3), pp. 255-262. Available at: https://uk.practicallaw.thomsonreuters.com/Document/I213DB4E0AD2811E9A588B8B29A362366/View/FullText.html (Accessed: 01 Sept 2019)
MrTvRetro (2012) History of Computer Animation. 24 February. Available at: https://youtu.be/tyvFerDZsuU (Accessed: 18 January 2020)
North, D. (2001) ‘Magic and Illusion in Early Cinema, Studies in French Cinema, 1:2, pp. 75
Nuffield Council on Bioethics (2018) Bioethics briefing notes: Artificial Intelligence (AI) in healthcare and research. Available at: https://www.nuffieldbioethics.org/assets/pdfs/Artificial-Intelligence-AI-in-healthcare-and-research.pdf (Accessed: 24 January 2020)
Ofcom (2019) Half of people now get their news from social media. Available at: https://www.ofcom.org.uk/about-ofcom/latest/features-and-news/half-of-people-get-news-from-social-media (Accessed: 28 January 2020)
O’Sullivan (2019) Lawmakers warn of ‘deepfake’ videos ahead of 2020 election. Available at: https://edition.cnn.com/2019/01/28/tech/deepfake-lawmakers/index.html (Accessed: 01 Sept 2019)
Porter, G., Kennedy, M. (2012) ‘Photographic Truth and Evidence’, Australian Journal of Forensic Sciences, 44(2), pp. 183-192. doi: 10.1080/00450618.2011.634835
Rogue One: A Star Wars Story (2016) Directed by G. Edwards [feature film]. Burbank, CA: Walt Disney Studios Motion Pictures.
Schwarts, O. (2018) You thought fake news was bad? Deep fakes are where truth goes to die. Available at: https://www.theguardian.com/technology/2018/nov/12/deep-fakes-fake-news-truth (Accessed: 01 Sept 2019)
The Vanishing Lady (1896) Directed by G. Méliès. [Short Film]. France: Georges Méliès.
VintageCG (2009) The Making of Brilliance (1985). 14 August. Available at: https://youtu.be/HZY5_ZzRdbk (Accessed: 26 January 2020)
Vox (2016) How Snapchat’s filters work. 28 June Available at: https://youtu.be/Pc2aJxnmzh0 (Accessed: 22 January 2020)
Wall Street Journal (2018) Deepfake Videos Are Getting Real and That’s a Problem | Moving Upstream. 15 Oct 2018. Available at: https://youtu.be/Ex83dhTn0IU (Accessed: 01 Sept 2019)
Warzel, C (2018) He Predicted The 2016 Fake News Crisis. Now He’s Worried About An Information Apocalypse. Available at: https://www.buzzfeednews.com/article/charliewarzel/the-terrifying-future-of-fake-news (Accessed: 01 Sept 2019)
Wired (2019) How Lyrebird Uses AI to Find its (Artificial) Voice. Available at: https://www.wired.com/brandlab/2018/10/lyrebird-uses-ai-find-artificial-voice/ (Accessed: 20 January 2020)
Woods, R. (2017). Ghostwatch: The BBC spoof that duped a nation. BBC News. [online] 30 Oct. Available at: https://www.bbc.co.uk/news/uk-england-41740176.
Young Sherlock Holmes (1985) Directed by B. Levinson [feature film]. Los Angeles, CA: Paramount Pictures.