As a new practice in the field of artificial intelligence, in recent years, the deep synthesis technology that uses deep learning, virtual reality and other generative synthesis algorithms to produce images, audio, video, virtual scenes and other information has been widely applied in many fields, and with the emerging demand for use, the number and attention of deep synthesis content have shown a rapid growth trend. But on the other hand, the audio and video generated by the malicious use of this technology has also shown great destructive power, causing reputation damage and property damage to individuals and enterprises, and also posing a threat to society and national security.
Recently, the Institute of Artificial Intelligence of Tsinghua University, Beijing Ruilai Intelligent Technology Co., Ltd., Tsinghua University Intellectual Media Research Center, National Industrial Information Security Development Research Center, and Beijing Big Data Center jointly released the "Ten Trends Report on Deep Synthesis (2022)" (hereinafter referred to as the "Report"), which comprehensively and deeply introduces and judges the opportunities and challenges brought by deep synthesis technology and application from many aspects such as technology research, field application and development trend, and gives practical suggestions and measures on its development and governance.
"In the thousands of years of human civilization history, the innovation of science and technology has brought about social changes across eras many times. From the fire of ancient times to the electricity of modern society, every technological change inevitably brings security risks while benefiting mankind," Zhang Cymbal, academician of the Chinese Academy of Sciences and founding dean of the Institute of Artificial Intelligence of Tsinghua University, wrote at the beginning of the "Ten Trends report on deep synthesis (2022)", "Unlike other technologies, because the security problems of artificial intelligence mainly stem from the insecurity of the algorithm itself, it is more likely to be misused and abused." Taking deep synthesis technology as an example, criminals can use it to pose a huge threat to society. ”
Landing applications in many fields, in-depth synthetic content to meet the explosive growth
Deeply synthesized images, videos, audio, text and other content, such as popular film and television drama clips, face-changing videos of topic stars, etc., have strong entertainment and dissemination, with the development and maturity of technology, more and more creators publish and share deeply synthesized content on the Internet, and the number is growing rapidly year by year.
Data description: In 10 domestic and foreign platforms (iQiyi, Tencent Video, Youku, Bilibili, Douyin, Kuaishou, Weibo, YouTube, Twitter, TikTok), 10 Chinese and English keywords such as "Deepfakes" were searched, and after deduplication through URLs, the data results were statistically counted
Taking video as an example, the Report shows that on mainstream audio and video websites and social media platforms at home and abroad, the number of newly released deeply synthesized videos in 2021 has increased by more than 10 times compared with 2017. In addition, the attention of deep synthetic content has also increased exponentially, and through the statistics of interactive data, the number of likes of newly released deep synthetic videos in 2021 has exceeded 300 million.
"The continuous maturity of technology is an important reason for the explosive growth of deep synthetic content. The continuous increase in research papers, the emergence of open source technology tools and a large number of representative methods have made the effect of deep synthetic content more realistic and more efficient, especially the emergence of algorithms such as generative adversarial networks, so that synthetic content has reached the level of 'true and false'. Tian Tian, CEO of Ruilai Wisdom, said.
Data description: In 10 domestic and foreign platforms (iQiyi, Tencent Video, Youku, Bilibili, Douyin, Kuaishou, Weibo, YouTube, Twitter, TikTok), 10 Chinese and English keywords such as "Deepfakes" were searched, and after the URL was deduplication, the data with video classification tags were statistically analyzed
Because of this, deep synthesis has developed diversified commercial applications in film and television production, advertising and marketing, social entertainment and other fields, such as AI synthesis anchors, virtual idols, repairing historical old photos, localized dubbing of film and television dramas, and "digital resurrection" deceased actors. Some synthetic products developed for the public have also been launched, and the "Report" shows that video, voice and text forms of services are the most common, such as special effects video production software that supports style customization; voice direction has derived applications such as voice navigation, audiobooks, and automatic news broadcasting; text synthesis plays an important role in news reporting, poetry creation, chat Q&A, etc.
The proposal of new business thinking such as "metaverse" also provides a broader application scenario for deep synthesis. "For example, virtual people and digital people are the main applications of deep synthesis and an important part of the 'meta-universe'," said Xue Hui, head of Alibaba's security perception and cognitive intelligence department. Chen Changfeng, executive vice dean of the School of Journalism and Communication at Tsinghua University, believes that deep synthesis will redefine the virtual digital space, and from the perspective of communication sociology, a new human survival scenario will be developed based on deep synthesis technology.
Risks have intensified, and technological testing has become an important response
Deep compositing inspires creativity in new content, but it also brings new threats. With the gradual "popularization" of technology, criminals can easily forge audio and video, carry out illegal acts such as framing, defaming, fraud, extortion, and even fabricate speech by state dignitaries to disrupt social and political order.
In October 2021, police in Hefei, Anhui Province, seized a case of illegally using deep synthesis technology to forge mobile phone users' faces and dynamic videos to crack identity verification, providing technical support such as registered virtual mobile phone cards for the black and gray industry. In recent years, similar incidents have begun to enter the public eye more.
Deep synthetic content blurs the boundary between true and false, and will have a huge impact on social trust, media trust, and political trust. Chen Changfeng said that the high difficulty of screening false content affects the effectiveness of fact verification, in the node of major social events or political events, deep synthesis technology may be used to manipulate public opinions, with the help of social media, so that false information in a short period of time to trigger a viral spread, intensify social contradictions.
As negative risks continue to intensify, how to effectively identify deep synthetic content has become the key, but with the continuous improvement of synthetic quality, traditional biometric-based identification methods are becoming more and more difficult to play a role.
Data description: In arXiv, 16 keywords such as "GAN" were searched within the scope of the abstract, and the number of papers published in the fields of image, video, audio, and text synthesis was counted
"At present, the detection of deep synthesis mainly relies on artificial intelligence models and the completeness of training data, including low detector versatility, applicability of public data sets, data sensitivity, etc., which will bring many challenges." Ren Kui, dean of the School of Cyberspace Security at Zhejiang University, said.
Wu Hequan, an academician of the Chinese Academy of Engineering, believes that there are two key points in the governance of deep synthesis, first, to continue to develop technology, can not be "one size fits all" prohibited, to avoid hindering positive application and innovation. Second, the derived security problems should be solved from the source, using technological innovation, technological confrontation and other ways to continuously improve and iterate the ability of detection technology. Tian Tian also said that the emergence of new forgery methods, the increasing complexity of the network communication environment, and the existence of vulnerabilities and defects based on detection algorithms, anti-deep pseudo detection technology faces "strong confrontation" and needs continuous update and iteration.
The "Report" shows that at present, academia and industry have invested a lot of research on anti-deep pseudo detection, meta (former Facebook), Google, Microsoft and other institutions have launched deep synthetic video certification methods or products.
In China, Tsinghua University, University of Science and Technology of China and other universities have achieved remarkable results in deep fake content detection, and deep forged content detection platform DeepReal launched by Tsinghua University incubation team Ruilai Wisdom has industrial-grade detection performance and detection ability to cope with changes in the real network environment.
Zhu Jun, director of the Basic Theory Research Center of the Institute of Artificial Intelligence of Tsinghua University, believes that deep synthesis detection faces continuous attack and defense and game, and in the future, it is necessary to integrate multi-modal content forensic analysis, traceability technology based on digital watermarking and other aspects of capabilities to achieve accurate identification.
Construct a multi-dimensional governance mechanism to guide the benign development of deep synthesis technology
The benign development of deep synthesis technology is inseparable from the exploration of multi-dimensional governance mechanisms.
The "Report" shows that in addition to the development of deep-depth forged content detection technology, in recent years, in response to the challenges brought about by the malicious use of deep synthesis technology, countries around the world have issued relevant laws and regulations to explore the governance path of deep synthesis. Internationally, the United States has made specific legislation at the federal and state levels, and the European Union has incorporated deep synthesis into existing legal frameworks such as the General Data Protection Regulation (GDPR). In addition, Germany, Singapore, the United Kingdom, South Korea and other countries have laws and regulations applicable to the trial of crimes related to deep synthesis technology.
The mainland is also actively exploring effective governance mechanisms. Since November 2019, the Provisions on the Administration of Online Audio and Video Information Services, the Provisions on the Ecological Governance of Online Information Content, the Civil Code of the People's Republic of China, and the Provisions on the Recommendation and Administration of Internet Information Service Algorithms have all put forward regulatory requirements of varying degrees for the generation of synthetic content. In January this year, the Cyberspace Administration of China (CAC) published the Provisions on the Administration of Deep Synthesis of Internet Information Services (Draft for Solicitation of Comments), which made specific provisions on the use, marking, scope of use, and penalties for abuse of deeply synthesized content.
For the exploration of the governance path of deep synthetic content, Chen Changfeng said that it can work hard from several aspects such as technology, ethics and legal system, "in terms of technology, through the participation of government and social organizations, collaborative governance, and ethics, establish and advocate relevant deep synthesis principles". Duan Weiwen, director of the Research Office of Philosophy of Science and Technology at the Institute of Philosophy of the Chinese Academy of Social Sciences, suggested that systematic and forward-looking interdisciplinary research should be carried out on social, legal and ethical issues caused by deep synthesis technology, and targeted governance and supervision should be carried out on high-risk application scenarios that may occur.
Zeng Yi, a researcher at the Institute of Automation of the Chinese Academy of Sciences and an expert of UNESCO's Ad Hoc Expert Group on AI Ethics, advocated the development of self-discipline and autonomy in industry and research, saying that before laws and regulations are fully mature and systematic, the industry itself should strengthen the awareness of "theory first" and jointly prevent abuse and strictly prohibit bad use as an industrial community.
"At the social level, we should increase publicity and popularization, strengthen citizens' understanding of artificial intelligence technologies such as deep synthesis, improve the awareness of prevention in the whole society, and promote citizens as responsible users of deep synthesis technology, actively identify synthetic content and actively practice social supervision," said Xu, an associate professor at the Law School of the University of International Business and Economics.
In this regard, Tian Tian also said that the essence of deep counterfeiting is the lack of transparency, so it is particularly important to improve the public's awareness of deep synthesis technology, only when the threshold is lowered to all audiences can recognize, discuss and understand this issue in a common framework, deep synthesis technology can develop healthily and benignly.
The "Report" also suggests that all parties should keep pace with the times to implement the new regulatory requirements, and under this premise, continue to pursue technological breakthroughs, continuously open up application scenarios for deep synthesis technology, create demonstration benchmarks, and form a driving effect on the artificial intelligence industry as a whole, so as to promote the sustainable and healthy development of new technologies.