In recent years, with advances in artificial intelligence technology, some criminals have begun using AI technology to merge other people's faces and voices, creating highly realistic synthetic images to carry out new types of online fraud. Such scams often cause significant losses to victims in a short period of time. How should we see through this kind of AI "deepfake"? Are there any good methods to identify "AI face-swapping"? Let's first look at a recent "AI face-swapping" fraud case that occurred in Xi'an, Shaanxi.

Zhang, a finance staff member in Xi'an, Shaanxi, had a video call with her boss, who asked her to transfer 1.86 million yuan to a designated account.

　　Victim Ms. Zhangsaid: The boss asked me to transfer the money quickly, and it was very urgent, because his voice and video image looked exactly like him, so I was even more convinced that this transfer was what he said, and then I transferred the money directly.

After the transfer, Ms. Zhang, in accordance with regulations, sent the electronic voucher to the company's internal finance group. Unexpectedly, when the boss in the group saw the message, he asked her about the reason for the funds.

　　Victim Ms. Zhangsaid: Then we called the boss again to verify, and the boss said he had not sent me a video and had not mentioned this transfer.

Realizing she had been scammed, Ms. Zhang quickly called the police for help. The police immediately coordinated with the anti-fraud center and contacted the relevant bank for an emergency stop payment, ultimately preserving most of the defrauded funds, 1.56 million yuan.

　　AI "multi-person face-swapping" fraud case: HK$200 million stolen

More than one million yuan in cash is a huge sum for most people. Fortunately, the police handled it in time and recovered most of the losses. Now let's look at another case, in which the "face-swapping" technology was more difficult and the amount involved was larger. Recently, Hong Kong police also disclosed an AI "multi-person face-swapping" fraud case involving as much as HK$200 million.

In that case, an employee of the Hong Kong branch of a multinational company was invited to attend a multi-person video conference initiated by the chief financial officer of the headquarters. As required, the employee made multiple transfers, transferring HK$200 million to five local bank accounts, and only learned of the fraud after checking with the headquarters. Police investigation revealed that in this case, in the so-called video conference, only the victim was a "real person"; all the other so-called participants were fraudsters after AI face-swapping.

After watching these two cases, you must be curious: what exactly is the technical principle behind AI face-swapping? At the technical level, how does it achieve precise recognition and replacement of faces and create realistic effects? Let's hear the expert's explanation.

　　How face recognition and replacement are technically achieved

　　Xue Zhihui, an expert from the Artificial Intelligence Security Governance Committee of the China Cyberspace Security Associationsaid: The AI face-swapping process mainly includes several key steps: face recognition and tracking, facial feature extraction, face transformation and fusion, background environment rendering, and image and audio synthesis. The core behind it includes three parts. First, deep learning algorithms are used to accurately identify the face image in the video and extract key facial features such as the eyes, nose, and mouth. Second, these features are matched, replaced, and fused with the target face image. Finally, through background environment rendering and adding synthesized sound, a highly realistic fake face-swapped video is generated.

　　Fast AI "face-swapping" can be done with just one photo

To understand how realistic AI face-swapping can be, the reporter, in cooperation with professional technicians, experienced AI face-swapping technology in depth.

The technician first used a mobile phone to take a photo of the reporter's face and imported it into AI software. To the reporter's surprise, although the person in front of the computer camera was the technician, the output was the reporter's photo. It was almost "one-click face-swapping," requiring no complex environment or decompression operations.

More surprisingly, as the technician's facial expressions changed, the reporter's face in the photo changed accordingly.

　　Reporter: Why does my photo move when the technician's face moves?

　　Xue Zhihui, an expert from the Artificial Intelligence Security Governance Committee of the China Cyberspace Security Associationsaid: First, through video capture, the face in the picture can be tracked and located. After locating it, the second step is to collect and extract facial feature points, mainly including obvious facial features related to the mouth, nose, and eyes. After collection, the third step is to transform, fuse, and reshape the original photo.

　　Reporter: What else can be done with this photo?

　　Xue Zhihui, an expert from the Artificial Intelligence Security Governance Committee of the China Cyberspace Security Associationsaid: At present, through this photo, the technician has already made this photo move and come to life. If this photo is further stored, and a large number of photos can be stored, then later this photo can be synthesized into a short video and released.

AI face detection technology is mainly realized through deep learning algorithms. This technology can identify facial features and analyze them precisely. It can extract a person's facial expressions from a photo or video and match them with another person's facial features. Experts told reporters that if AI face-swapping technology is to be used in real-time video calls, one photo is far from enough; it would require the collection of nearly a thousand photos from different angles.

　　Xue Zhihui, an expert from the Artificial Intelligence Security Governance Committee of the China Cyberspace Security Associationsaid: If real-time point-to-point communication is required, more photos need to be collected, and then deep learning algorithm model training is carried out. After the model is trained, it can be fed into the video to achieve real-time changes and conversion. In this case, real-time face-changing can be achieved. Voice communication is also possible; some voices of the target person need to be collected in advance, and then model training is carried out to restore the target person's voice.

　　AI-generated highly realistic videos: high difficulty and large investment

Experts introduced that for fraud purposes, implementing point-to-point video calls requires AI to generate highly realistic videos. To achieve an effect realistic enough to be used for fraud is not easy.

Fang Yu, a member of the Digital Economy and Security Working Group of the Security Professional Committee of the China Computer Federation: For fraud, it is mainly through point-to-point video calls. At this time, if face-swapping technology and voice synthesis are used to carry out real-time face-swapping fraud, completing these technical operations requires strong technical support.

　　Xue Zhihui, an expert from the Artificial Intelligence Security Governance Committee of the China Cyberspace Security Associationsaid: Behind it there needs to be a large amount of capital investment, including photo collection, professional algorithm personnel, etc., and a long cycle is needed, including investment in computing power and algorithms in various aspects. Long-term continuous iteration and operation are needed to achieve a fairly realistic effect and possibly realize the actual effect of fraud.

　　At present, AI "face-swapping" is more used in entertainment video recording

In addition to some criminals attempting to use AI technology to commit fraud, in fact, in recent years, AI technology has been more used in secondary creation of short videos and is entertainment in nature. More and more celebrity AI face-swapped videos are appearing on the internet, and many netizens sigh that such videos "match in mouth shape and gestures, so natural that I almost thought it was real."

For example, this software, after taking a photo of the reporter's face, can record and generate a video in which the reporter instantly becomes a racing driver.

Fang Yu, a member of the Digital Economy and Security Working Group of the Security Professional Committee of the China Computer Federation: What we often see now in AI technology is mainly short-video face-swapping, through doing some specific actions, dancing, and so on. These videos actually look somewhat unnatural and are purely entertainment-oriented face-swapping.

The reporter found that in mobile app stores, there are dozens of face-swapping software programs that can achieve the purpose of face-swapping.

　　Xue Zhihui, an expert from the Artificial Intelligence Security Governance Committee of the China Cyberspace Security Associationsaid: From the perspective of entertaining the public, there are now many software and tools on the market that can achieve the effect of AI face-swapping, but the realism is only about six or seven out of ten, and the public can see it directly. But if a fraud video is to be generated, it is necessary to generate point-to-point videos with extremely high realism.

　　AI technology "face-swapping and voice-swapping" may involve legal risks

However, AI technology is also a double-edged sword. Even for entertainment, using AI face-swapping and AI voice-swapping carries legal risks. Legal experts said that using AI technology to swap faces and voices for others, or even translate into other languages and publish videos, may be suspected of infringement, mainly in three aspects:

First, it may infringe copyright. For example, crosstalk, sketches, and other works are all "works" protected by the Copyright Law of the People's Republic of China. For example, if netizens use AI software to "translate" crosstalk, sketches, and other works into other languages, authorization from the copyright owner is required; otherwise, there is an infringement issue.

Second, it may infringe portrait rights. According to the Civil Code of the People's Republic of China, no organization or individual may infringe another person's portrait rights by means of vilification, defacement, or forgery using information technology. Without the consent of the portrait rights holder, no one may produce, use, or publicly disclose the portrait of the portrait rights holder, except as otherwise provided by law.

Third, it may infringe voice rights. According to the Civil Code of the People's Republic of China, the protection of a natural person's voice shall refer to the relevant provisions on the protection of portrait rights. That is to say, the consent of the voice rights holder is required to use another person's voice.

　　Learn a few tricks to easily identify AI "face-swapping and voice-swapping"

The emergence of AI face-swapping technology means that hearing may be false and seeing may not necessarily be believing. So how should we prevent it? Experts said that no matter how realistic AI face-swapping is, there are still some methods to identify whether a video is real or fake.

　　Fang Yu, a member of the Digital Economy and Security Working Group of the Security Professional Committee of the China Computer Federationsaid: In fact, from the real-time deepfake videos I have seen so far, it is still possible to verify them through some methods.

For example, we can ask the other party to wave their hand in front of their face during a video conversation for identification. Real-time forged videos require real-time generation and processing and AI face-swapping.

Then, during the waving process, interference with facial data will be caused, and the final effect is what we see: during waving, the forged face will produce a certain jitter or some flickering, or some abnormal situations.

The second is that in point-to-point communication, you can ask some questions that only the other party knows to verify the other party's authenticity.

　　Raise awareness of prevention and avoid leakage of personal biometric information

Experts said that in addition to some tips for identifying AI face-swapping fraud, each of us should raise our awareness of prevention and also take relevant preventive measures in daily life and develop good internet habits. First, we should do a good job in daily information security protection and strengthen the security protection of biometric data such as faces, voiceprints, and fingerprints; in addition, do a good job in the security management of software and hardware of personal terminal devices such as mobile phones and computers. Second, do not log in to websites of unknown origin to avoid virus intrusion. Third, for applications that may collect voice, images, and even video and location, do a good job in authorization management. Not giving others the opportunity to collect your information can also, to a certain extent, keep AI face-swapping fraud away from you.

　　With the development of AI technology, multi-level supervision and regulation are needed

In addition to raising self-protection awareness, how to strengthen supervision of AI technology has also become an issue of increasing concern to more and more people. AI technology itself is not the problem; the key is how we use it and how to form effective supervision. Experts introduced that the development of AI technology requires multi-level supervision and regulation.

First, at the source, it is necessary to further strengthen the protection of citizens' personal information, especially to strengthen technical and judicial protection of private information such as biometric features. Second, management can be strengthened at the technical level. For example, video dissemination websites or social software can be required to use professional identification software to verify and mark AI-generated videos with an irremovable "AI-generated" watermark. At present, this digital watermark authentication technology needs to be further popularized. Third, at the legal and regulatory level, relevant laws and regulations in fields such as artificial intelligence should be further improved. On August 15, 2023, China officially implemented the Interim Measures for the Management of Generative Artificial Intelligence Services. The Measures "draw red lines" in multiple aspects, aiming to promote the healthy development and standardized application of generative artificial intelligence.