深度伪造技术对信息战的影响（英）-24页

July 2022PerspectiveEXPERT INSIGHTS ON A TIMELY POLICY ISSUETODD C. HELMUSArtificial Intelligence, Deepfakes,and DisinformationA PrimerDisinformation is getting an upgrade. A primary tool of disinformation war-fare has been the simple meme: an image, a video, or text shared on socialmedia that conveys a particular thought or feeling (Sprout Social, undated).Russia used memes to target the 2016 U.S. election (DiResta et al., 2019);China used memes to target protesters in Hong Kong (Wong, Shepherd, andLiu, 2019); and those seeking to question the efficacy of vaccines for coronavirusdisease 2019 used memes as a favorite tool (Wasike, 2022; Helmus et al., 2020).By many accounts, memes, as well as other common and seemingly old-fashioneddisinformation tools such as fake news webpages and stories and strident Facebookposts have successfully undermined confidence in U.S. elections (Atlantic Coun-cils Digital Forensic Research Lab, 2021), sown division in the American electorate(Posard et al., 2020), and increased the adoption of conspiracy theories (Center forCountering Digital Hate, 2021; Marcellino et al., 2021). Advances in computer sci-ence and artificial intelligence (AI), however, have brought to life a new and highlycompelling method for conveying disinformation: deepfakes. Deepfake videos areCORPORATION Deepfake VideosFIGURE 1A Still Image from a TikTok Video Producedby deeptomcruiseAs previously noted, deepfake videos include syntheticallymodified footage that presents alterations in subjects facesor bodies. These synthetic videos images are developedthrough generative adversarial networks (GANs). Tianx-iang Shen, Ruixian Liu, Ju Bai, and Zheng Li (2018) providean excellent description of how GANs work to create syn-thetic content:The GAN system consists of a generator that gener-ates images from random noises and a discriminatorthat judges whether an input image is authentic orproduced by the generator. The two components arefunctionally adversarial, and they play two adver-sarial roles like a forger and a detective literally. Afterthe training period, the generator can produce fakeimages with high fidelity. (p. 2)Since Ian Goodfellow and colleagues created the GANsystem in 2014 (Goodfellow et al., 2014), deepfake videoshave become increasingly convincing. In spring 2021, aTikTok account (Tom deeptomcruise, 2021) released aseries of highly realistic deepfake videos of what appearedto be Tom Cruise speaking. As of that time, the video hadmore than 15.9 million views and has spurred significantpublic angst about the coming age of deepfake disinforma-tion (see Figure 1).SOURCE: Tom deeptomcruise, “Sports!” 2021.NOTE: As of April 12, 2022, this TikTok video had more than 16.1 million views.Well-crafted deepfakes require high-end computingresources, time, money, and skill. The deepfakes fromdeeptomcruise, for example, required input of manyhours of authentic Tom Cruise footage to train AI models,and the training itself took two months. The deepfakes alsorequired a pair of NVIDIA RTX 8000 graphics processingunits (GPUs), which cost upward of US$5,795 each (as ofthis writing). The developers then had to review the finalfootage frame by frame for noticeable tells, such as awk-ward or non-lifelike eye movements. Finally, this processcould not have happened without a talented actor whocould successfully mimic the movements and mannerismsof Tom Cruise (Victor, 2021; Vincent, 2021).3 Over time, such videos will become cheaper to createand require less training footage. The Tom Cruise deep-fakes came on the heels of a series of deepfake videos thatfeatured, for example, a 2018 deepfake of Barack Obamausing profanity (Vincent, 2018) and a 2020 deepfake of aRichard Nixon speecha speech Nixon never gave (MITOpen Learning, 2020). With each passing iteration, thequality of the videos becomes increasingly lifelike, and thesynthetic components are more difficult to detect with thenaked eye.call from someone who sounded like his boss at a parentcompany. At the instruction of the voice on the phone,which was allegedly the output of voice-cloning software,the CEO executed a wire transfer 220,000 (approximatelyUS$243,000) to the bank account of a Hungarian supplier(Stupp, 2019). In another example, a Philadelphia manalleged that he was the victim of a voice-cloning attack; hewired US$9,000 to a stranger when he believed he heardthe voice of his son claiming that he was in jail and neededmoney for a lawyer (Rushing, 2020).Various webpages now offer access to deepfake ser-vices (see Meenu EG, 2021). Popular sites include Reface(undated), which allows users to swap faces with faces inexisting videos and GIFs; MyHeritage (undated), whichanimates photos of deceas