What is deepfake and why is it a serious threat?
Deepfake – a technique that allows to synthesize human images based on artificial intelligence often used to create, edit, distort, cut together, and record video clips in a sophisticated way – has been posed challenges and threats to the integrity of online information. Along with the increasing accessibility of deepfake technology and the speed of internet platforms, deepfakes are rapidly gaining popularity and reaching millions of people.
In addition to the proposed technological solutions to detect and prevent the spread of deepfakes, each social network user needs to be conscious of protecting his or her personal information.
About Deepfake
With the strong development of information technology, today social networks have spread and covered the world. Besides the great benefits that social networks bring, we are facing many risks and challenges, even threatening national security and social order and safety. In particular, it must be mentioned the negative effects from bad and toxic information spread on social networks as well as the problem of fake news – Fake News. About 30 years ago, the appearance of Photoshop image editing software completely changed the way people received news when the images could completely be the product of a feat. Internet users began to doubt the accuracy of images, still placing more trust in video and audio recordings because these are almost impossible to fake. But again, Deepfake appears to allow creating and editing, distorting, cutting clips, and recording in a sophisticated way that has been posing challenges and threats to the integrity of information. online. Experts say that: Deepfake, when abused to create fake information, will cause unpredictable consequences. Anyone who publishes their personal profile on social media can be spoofed and become a victim of deepfake.
Deepfake [1],[2] (a combination of “deep learning” – deep learning and “fake” – artificial in English that has appeared publicly on the Internet since the end of 2017) is a technique that allows total Synthesize human images based on artificial intelligence. It is used to combine and superimpose source images and videos using machine learning techniques known as general adversarial networks. In other words, DeepFake is a technique of creating videos, images, voices, and sounds with artificial intelligence software that edits, distorts, and cuts them to make them look real.
Back in 2013, when Paul Walker, the actor of the Fast and Furious series, suddenly passed away while the 7th episode still hadn’t finished filming. The crew invited Cody Walker, Paul’s younger brother, to act as a stunt double because he had similar facial features so they could easily make Cody look like his brother in post-production. On the day of its release, the film made millions of “fans” around the world surprised by its authenticity. This is the most primitive form of Deepfake.
The technology became widely known when a user on social network Reddit used it to swap celebrity faces into adult movies with fairly high accuracy without much effort. However, these tools have been growing in popularity for about two years now, threatening not only stars but ordinary people as well. Many famous figures such as former US president Barack Obama, former first lady Michelle Obama, Facebook founder Mark Zuckerberg, President Donald Trump and Speaker of Congress Nancy Pelosi have all been victims of DeepFake. Lawmakers and digital rights organizations also predict DeepFake will have a big impact on the 2020 US presidential election.
How to create Deepfake
First, the original video is scanned in and decomposed into phonemes (the sounds that make up words) spoken by the “victim” (who will be spoofed). These phonemes are connected to the corresponding facial expressions when expressing each of them. Next, a 3D model of the lower half of the speaker’s face was created using the original video. When editing a video’s textual content, the software combines all this collected data including phonemes, images, and 3D face models to create new footage that matches the added content. into text. Then, the software will paste the finished product into the original video to create the final video. [2]
In the early stages, Deepfake technology is still not really perfect because the algorithms only work on videos of the face of the speaker and need about 40 minutes of input data, the sound quality is quite low compared to the original. , it is difficult to change the mood or tone of the speaker. In addition, if the speaker’s face is obscured whenever, like someone waving while speaking, the algorithm will fail completely. But these limitations have almost been overcome. In the latest example of new Deepfake technology published at the end of June 2019, researchers have shown that new software using machine learning algorithms can allow users to edit written content. transcribe the video to add, remove, or change the words that come out of the speaker’s mouth in the video.
The contribution of artificial intelligence used in deepfake technology basically goes through two steps: Collecting training data and learning the model through the selection of algorithms for continuous processing, learn from the training data. In the case of deepfakes, the input training data are pre-existing videos and images of the “victim”, the learning algorithm is now publicly available in the open source form of TensorFlow or Keras. After a period of “learning”, the machine was able to perform a face-matching task with high similarity.
Besides the problem of the abundance of data and algorithms already available on the network, more worrying is that the process of learning and processing also does not need supercomputers. Artificial intelligence researcher Alex Champandard thinks that a computer with a popular graphics card can do the above process in just a few hours. Without a graphics card, the central processing unit (CPU) can still perform tasks for longer, up to several days.
The creation of fake videos in a sophisticated but not too complicated way like the analysis above proves the danger of deepfake technology not only to the lives of each “netizen” in general, especially, especially are famous people, politicians in particular, but also threaten national security, economic and political life of a country and need to have solutions to this problem.
Deepfake’s Malicious Apps
The first deepfake video appeared in 2017, in which a celebrity’s face was swapped into that of a porn actor. Since then, the technology has been used to influence public opinion. Journalist Rana Ayyub was attacked by deepfakes when her face was turned into a pornographic video as if she had acted in it. This malicious attack happened shortly after she campaigned for justice for rape victim Kathua. In addition, Deepfakes can also serve malicious personal motives. A mother from Pennsylvania allegedly used deepfake videos to try to get her teenage daughter’s cheerleaders off the team.
The case of Malaysian Economy Minister Azmin Ali is an interesting case to consider. In 2019, Azmin was involved in a sex tape scandal in which he allegedly had an affair with a member of his political party, Haziq Abdul Aziz. Since “sodomy” (the term deviant sex) is a crime in Malaysia that carries a prison sentence of up to 20 years, Azmin tried to evade the consequences by claiming that the video was a scam. deep. Although no one has been able to definitively determine it, public opinion and digital forensic experts are still divided into two camps. Since the public cannot distinguish what information is true from false, they are more willing to accept their preconceived biases.
As a result, this has had serious consequences for the reputations and political careers of the parties involved. In addition, this has led to the destabilization of the governing coalition and the strong rejection of LGBT rights in Malaysia. According to Towes, deepfakes not only worsen social divisions and disrupt democratic discourse, but also undermine public safety and undermine journalism.
Deepfake detection technology solutions
Due to the characteristics of the online media age, the image quality can be very low and not need specific context but still reach the viewer. Like fake news, a lot of fake news content can be easily debunked with just a few minutes of a Google search, but that doesn’t stop them from spreading. Deepfake too, the products of this technology do not need to be perfect in every detail and still have a lot of people to accept. Besides, preventing deepfake encounters many difficulties due to the anonymity of the forums, legal experts say that it is very difficult to find the attacker; Limiting by tools is also not simple. Meanwhile, the data is the pictures that users make publicly available online and the open source algorithms provided by the companies to researchers, students and everyone else interested in machine learning technology. .
Based on how fake videos are created, there are currently two proposed solutions to prevent deepfake: using high-precision Deepfake detection tools and news photo and video authentication mechanisms. .
Solution 1: Develop a deepfake video detection tool
In August 2018, researchers at the US Department of Defense Research Center (DARPA) introduced a tool that allows the identification of deepfake videos [3]. This tool can also analyze details that are not visible to the human eye, such as analyzing the spectrum or light of an image to recognize distinct locations. However, the scientists who invented the tool admit they are still evolving to keep up with the latest counterfeiting techniques.
Deepfake [4] uses neural networks to replace the face of the person in the original video with the face of the object you want to fake. Although the neural network does this task very efficiently, it cannot understand the physical and natural features of a human face. Therefore, deepfakes can be detected through a number of unnatural behaviors – the most notable of which are not blinking, head movements.
Researchers at the University of California, Riverside, have developed an AI model that can detect fake images by focusing on the edges of objects appearing in an image. Usually, the boundary between the original image and the merged objects will have some unique characteristics, such as unnatural smoothness or blur. This model is trained using a large dataset labeled between edited and unedited images. Thereby, the neural network will be able to find pattern rules, thereby distinguishing the edges of edited objects in the image, even with new images that do not belong to the dataset. While this model has only been used on still images, it will soon be available on video as well. Basically, deepfake videos are just a series of images that run continuously so they can be detected using the same mechanism.
However, according to current assessment, Deepfake and anti-Deepfake technology is developing in the same direction as the relationship between viruses and anti-virus software. Any bugs or weaknesses discovered can make anti-virus more effective, later versions of the virus will “patch” these errors and become even more sophisticated than before. As in the case of Deepfake, early versions can cause the character’s edited face to always be in an open state, or the mouth when pronouncing it is not natural. Subsequent versions have enhanced and improved these elements so markedly that it is difficult to tell the difference between the original and the edited version with the naked eye. Even as technology continues to advance, detection techniques, often lag behind the most advanced methods of deepfake video creation.
Solution 2: Video and photo authentication solution
Currently, most deepfake detection efforts are focused on finding evidence of tampering, ignoring another very effective method: proving what is true. This is the same method used in the Archangel project, which was carried out by researchers at the University of Surrey, UK, and is currently being tested on multiple national repositories.
Archangel uses a combination of neural networks and blockchain, to create a smart video storage for later use during authentication. After a video is added to the archive, Archangel trains the neural network to use various formats of this video. From there, the neural network will be able to detect if the new video is compatible with the original stored video, or has undergone editing.
In traditional comparison techniques, files will be authenticated on a byte-by-byte basis – a method that is not suitable for video because the structural characteristics change across the format. However, Archangel’s neural network compares by pixels, using codec-agnostic decoding.
To ensure that the neurons are not manipulated, Archangel houses them on a blockchain maintained by the government archives involved in the test of this project. As such, adding data to the archive will need the consent of the parties involved. Therefore, no organization can decide for themselves which videos are real and which are fake. It is expected that, after Archangel is made public, everyone will be able to try to authenticate their videos on this program.
One obvious weakness of this approach is that each video requires a differently trained neural network – which is extremely costly in terms of time and computational power required. This method is especially suitable for videos with sensitive content such as recordings of Parliament and speeches of politicians.
Solution 3: Solution for social network users
Deepfake is becoming more and more sophisticated and anyone can fall victim to Deepfake, especially women, celebrities and politicians. While the legal protection mechanisms really have not gone anywhere, it is extremely difficult to find out who the attacker is. protect yourself.
When deciding to join a social network or share personal information online, users must understand the following information and mechanisms: Who can access the information they post online? Who controls and owns the information we post on social media? What information about myself and my contacts will be passed on to others? Do my communication partners share their information with others? Do I trust people to be connected on social media?
Here are some solutions to help protect users on social networks:
Use secure passwords to access social networking sites and make it a habit to change your passwords periodically.
Understand the default settings in the privacy settings of a social networking site and how to change them.
Consider using separate accounts/identities, or possibly using fake names for different campaigns and activities. Since the key to using a secure network is the ability to trust network members, segregation of accounts can be a good way to ensure that this trust is achievable.
Be wary of accessing social media accounts at public internet hotspots. Delete login history and passwords after using a public machine.
Use the security protocol https when accessing social networking sites to protect usernames, passwords and postings.
Beware of putting too much information on your status updates and set up privacy settings that allow who can see what you share.
Be wary of integrating social networking sites because it is possible to remain anonymous on one site but reveal your identity on another.
Posting personal information: Social networking sites often require users to enter a lot of information about themselves to facilitate other members to find and connect with you. Perhaps the greatest risk posed to users of these sites is the possibility of identity spoofing, which is quite common.
Friends, readers, and communication partners: Consider only making friends with people you know well and trust that won’t use the information you post for malicious purposes.
Consider using geolocation – GPS, many applications can provide information about the user’s location.
Share photos/videos: Be aware and consider sharing photos or videos that can easily reveal people’s identities.
Online messaging: Consider using built-in online messaging apps in social networks (using the secure protocol https://).
Join/create groups, events, and communities: Consider joining and sharing personal information on groups, forums, and events.
Conclude
Deepfake is becoming more and more sophisticated and has a wide influence not only for a small individual or organization, but even for large enterprises, creating profound upheavals, causing military agitation or influence. influence politics. In Vietnam, hostile and anti-regime forces always find ways to distort, defame, and spread fake images and videos to falsify the truth. Fighting deepfake is not easy. While the giants in the technology industry are rushing to find solutions to detect and reduce the harms caused by deepfakes, users need to be wise in protecting themselves as well as identifying information online. society. Regulatory agencies need legislation to detect, prevent and handle these sophisticated counterfeiting acts in order to protect digital sovereignty in cyberspace and protect people from painful “problems” in cyberspace. digital era.
References:
[1] Hiểm hoạ của tin giả “bình dân” – Cheapfake. (2021), Bản tin an toàn không gian mạng, Bảo vệ chủ quyền số quốc gia, Cục An toàn thông tin.
[2] Westerlund, Mika. (2019) Tổng quan về sự bùng nổ của công nghệ deepfake. Tạp chí quản lý đổi mới công nghệ 9.11
[3] Cẩm Yến, Internet và cơn ác mộng mang tên Deepfake, số 29 (2019), Báo Thế giới và Việt Nam
[4] Yu P, Xia Z, Fei J, Lu Y. (2021) Khảo sát về kỹ thuật phát hiện giả mạo video deepfake. Tạp chí sinh trắc học – Viện kỹ thuật và công nghệ IEF
[5] Julie Posetti, Cherilyn Ireton, Claire Wardle, Hossein Derakhshan, Alice Matthews, Magda Abu-Fadil, Tom Trewinnard, Fergus Bell, Alexios Mantzarlis. (2018), Báo chí, tin giả và tin xuyên tạc, Sổ tay Unesco về giáo dục và đào tạo báo chí
[6] Tổng luận Khoa học – Công nghệ – Kinh tế (3/2021), Quản trị công nghệ của cách mạng công nghiệp lần thứ tư. Cục thông tin khoa học và công nghệ quốc gia.
