ChinAI #88: Deepfake Drama
The First Appearance of AI Face-Swapping on a Chinese Web TV Series
Welcome to the ChinAI Newsletter!
Greetings from a land where they promised us hover cars and we got deepfake porn instead…
…as always, the archive of all past issues is here and please please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay support access for all AND compensation for awesome ChinAI contributors).
Feature Translation: First Appearance of AI Face-Swapping in Chinese Web Series
If you want a pretty clear example of how society shapes technology just as much if not more than technology shapes society, look no further than deepfakes. This week we look at the face-swapping special effects of the web drama “Love of Thousand Years” (三千鸦杀), which did not work well (distorted faces, stiff expressions, dissonance between neck and head). It’s no Tiger King but it did invite a recent wave of ridicule comparing the face-swapping effects to a horror movie, which ultimately led more people to watch the show just to see the face-swap (which means it kinda worked out for the show in the end, I guess).
Context: After filming the drama, one of the actresses (Liu Lu) was involved in a kerfuffle with public transport authorities regarding the allowable amount of flammable compressed gas cans. She was reportedly blacklisted on a “bad artists list” [劣迹艺人] and Mango TV terminated their contract with her. To avoid the new drama from being suppressed, the producer swapped her face with that of another actress. Sourced from a longstanding ChinAI favorite, jiqizhineng (Synced), this week’s article digs deeper into this “first” case of deepfakes used in TV and film.
Key Nuggets:
Traces history of DeepFakes back to 2017 when a Reddit user face-swapped Gal Gadot’s face onto a mature film. References important developments in China, such as in February 2019 when a video swapping the face of one of China’s best-known actors, Yang Mi, into a classic Hong Kong TV drama, The Legend Of The Condor Heroes, went viral, picking up an estimated 240m views before it was removed by Chinese authorities (see related Guardian article). Subsequently, ZAO, a face-changing app also got really hot in China.
Face-Swap Black Market Industries: a lowered threshold for face-swapping has gradually created an industrial chain where face-swap software and technology are provided upstream, video and photo customization is supplied midstream, and the finished erotic videos are sold downstream. “Relevant products are available for sale in Tieba, QQ groups…the prices of finished erotic videos range from 2 RMB for 1; to 30 RMB for 46; 100 RMB for 150; and 100 RMB for 200, etc. Generally, they are sold in packages, and the videos mainly feature domestic first- and second-tier female stars.”
The piece notes the possible application of China’s May 2019 data security measures to govern DeepFake-altered videos: In China, at the end of May 2019, the Cyberspace Administration of China and relevant departments issued the ‘Data Security Management Measures (Draft for Comment),’ which requires that network operators who use big data, artificial intelligence and other technologies to automatically synthesize news, blog posts, forum posts, comments etc., to clearly mark such information as “synthesized”; they should not automatically synthesize information with the aim of seeking benefits or harming other people’s interests.” Note: It seems like whenever I come across a relevant regulation for a particular piece of Chinese tech news, DigiChina has already done the translation. Check out their excellent translation of the draft data security management measures, which I drew from for this section, here.
Still large obstacles to high-performance applications of face-swaps in TV and film: not something ordinary folk can do as it requires a relatively high-quality PC, a very good graphics card, and a decent amount of server usage. Problems related to personal privacy of celebrities and black market industry issues are being magnified as well.
Go Deeper:
Interested in the difference between various face-swap architectures? The piece references DeepFakes, which are the “classic” face-swap; Face2Face: focuses on “facial expression” manipulation -- examples include high-quality videos of a person (e.g. Obama) changing what he is really saying in a target video; and CycleGAN: unpaired image-to-image translation using a generative adversarial network model architecture (e.g. translating a summer landscape to winter rather than translating X specific person’s face to another Y specific person)
Want to get in the weeds of the economics of GAN-powered face-swaps vs. “manual” face-swaps (i.e. going frame by frame and manually editing, I assume?) Market price for manual face-swaps is a couple 100,000 RMB per minute, whereas the price of AI face-swapping is about 15,000 RMB per minute.
FULL(ish) TRANSLATION: First Appearance of AI Face-Swapping in Chinese Web Series
Must-read: The Precipice: Existential Risk and the Future of Humanity
At the beginning of this month, I was whining about our narrow view of national security, and recommended this book by Toby Ord, a Senior Research Fellow at the Future of Humanity Institute. I feel like the events of this past month have inspired many of us to have considerably widened our views. This book, now available in the US & Canada, presents a grand vision of the potential human flourishing of the future as well as a wake-up call to the existential catastrophes (e.g. climate change, engineered pathogens, nuclear weapons, unaligned artificial intelligence) from which we could never come back. In combination, ending these existential risks is among the most pressing moral issues of our time. The link lets you subscribe to Toby’s newsletter to download the first chapter now.
Should-read: OpenMined for Maximizing Privacy and Effectiveness in COVID-19 Apps
There is a boom in apps related to COVID-19 occurring right now, the team at OpenMined is trying to help those who are building/auditing/procuring such apps to do so in a way that helps reduce economic and epidemic threats to society while also checking against the erosion of privacy in the process. See their live document here. OpenMined is a project that aims to "decentralize AI,” led by Andrew Trask, a PhD Student at the University of Oxford and a research affiliate at GovAI.
Should-read: ChinAI Syllabus
With Sophie-Charlotte Fischer, Brian Tse, and Chris Byrd, we compiled a preliminary syllabus of readings on China’s AI landscape, which covers a range of topics. Inspired by Remco Zwetsloot’s really useful syllabus on AI and International Security. Grateful to Emmie Hine for her help and proposing the tutorial that sparked the syllabus in the first place as well as others, esp. Jade Leung, for suggestions. Hope is for this to be a living document so please send recommendations for stuff to be added.
Should-read: Megvii Open-Sources Deep Learning Framework
For SCMP, Sarah Dai and Minghe Hu in Beijing report on Megvii’s open sourcing of its deep learning framework (MegEngine). Dai and Hu get some really interesting quotes from Gao Wen, a Peking University professor who is also director-general of the country’s New Generation AI Technology Innovation and Strategic Alliance: “The latest tide of AI cannot live without deep learning technologies, whose development has everything to do with open-sourced infrastructures.” The piece cites Gao saying that Tensor Flow and PyTorch hold a combined 95 per cent of market share for deep learning frameworks. The framework, which contains 300,000 lines of code in its alpha version, is available for downloads at the company’s website and Github.
Thank you for reading and engaging.
These are Jeff Ding's (sometimes) weekly translations of Chinese-language musings on AI and related topics. Jeff is a PhD candidate in International Relations at the University of Oxford and a researcher at the Center for the Governance of AI at Oxford’s Future of Humanity Institute.
Check out the archive of all past issues here & please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay for a subscription will support access for all).
Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99