The deepfake maker shows how to identify a fake face
(Dan Tri) - With just one photo, AI can now create a deepfake face that can move. However, some details can help viewers detect abnormalities.
Just one photo can create a fake face
A face appears on the screen, blinks eyes, moves lips, changes expression according to who is speaking. If you only look for a few seconds, the viewer will not always realize that the image has been intervened with artificial intelligence (AI).
According to Mr. Vu Duy Hung, founder of Color Mind, a unit that applies AI to content and cinema production, who has many years of working with AI facial reconstruction technologies, at a simple level, the process of creating a deepfake face can start from just a photo.
"A deepfake face, in the simplest way, only needs a photo and we can completely do it," Mr. Hung said.
Mr. Vu Duy Hung said that from just one photo, AI can now create a deepfake face at a basic level (Photo: Tien Dat).
However, a photo does not mean that AI has complete information about that person's face.
If there is only a frontal photo, parts that do not appear in the original data such as side profile, side of the face or some expressions will have to be predicted by AI.
According to Mr. Hung, the more images and videos there are at different angles, the more data the model has to learn facial features.
Public videos on social networks can also provide many consecutive frames, recording how a person smiles, talks, turns their head or changes expression.
This is also the reason he believes that publishing too much personal data on the Internet can create additional risks as deepfake technology becomes increasingly accessible.
How does AI “learn” a face?
According to Mr. Hung, to recreate a person with high accuracy, data is the first important factor.
In the facial reconstruction projects he has performed, the team often collects images from many facial angles, videos to record natural gestures, and in some cases audio data to record voice quality.
Once the data is available, not all images go straight into the AI.
The operator must select data of good quality, while eliminating images that may cause noise.
"If we put noisy data in, the AI will learn incorrectly and cannot guarantee good results," Mr. Hung explained.
To create a more realistic deepfake, the AI model needs more data about many facial angles, expressions and movements of the character (Photo: Tien Dat).
The selected portion of data is then fed into the model for repeated training.
Through this process, the system gradually learns facial features, from the overall shape to the details and expressions that appear in the input data.
The trained model can be applied to another video source to create an image of the person's face that needs to be recreated.
After being selected, the data will be fed into the model for training, thereby creating a face that can be applied to videos or other images (Photo: Tien Dat).
However, AI does not just "copy" what is available. When data is missing, the model can interpolate to supplement it.
For example, if there is only data on one side of the face, the AI can rely on the assumption of symmetry to propose using information from the other side to create the missing part.
This makes reconstruction possible when documentation is limited, but also creates a gap between real data and what the AI predicts itself.
Why is deepfake still sometimes "exposed"?
According to Mr. Hung, one of the important technical issues when reconstructing a face is the ability to track facial features.
AI needs to recognize and follow certain locations to know how the eyes, nose, mouth or facial contours are changing in each frame.
When the character turns to an angle that causes too many points to be obscured or disappeared, the model may have difficulty continuing to adhere to the face.
As a result, images may appear distorted, shaky, distorted or unnatural.
Deepfake can still be revealed through eyes, mouth, teeth or unnatural expressions, although these errors are increasingly being corrected by AI (Photo: Tien Dat).
This is also the reason why experienced deepfake makers often have to calculate from the filming stage, choosing a camera angle that suits the capabilities of technology at the present time.
However, according to Mr. Hung, the most difficult part is still the expression and charisma of a person.
"The most difficult point is the data about the character's expressions, emotions and charisma movements. This is still one of the parts that requires a lot of post-production," he said.
A face may have the correct shape, but the eyes, smile or very small movements can still make the viewer feel "something isn't real".
Eyes, mouth and teeth were once easily recognizable features
From direct experience in making deepfake, Mr. Hung believes that currently there are still some details that can help viewers detect abnormalities.
First is the eyes.
In some videos, the deepfake character's eyes may move unnaturally or may not show the same charisma as a real person.
Next is the mouth area. The mouthpiece sometimes lags behind motion or sound.
In particular, teeth are one of the details that are difficult to reproduce properly if the product does not go through a thorough post-production process.
Another sign is the skin.
According to Mr. Hung, some AI-created characters still have the "plastic skin" phenomenon, meaning the skin surface is too smooth and uniform, creating a feeling more like artificial material than real skin.
An AI video on Instagram was created by a content creator in the style of a 1980s movie so realistic that many people mistakenly thought it was an excerpt from an old movie (Photo: Tien Dat).
However, he noted that the above signs are not a formula that can be used forever.
Technology is improving rapidly, and many errors that were once very noticeable in previous AI products have gradually been fixed.
"Nowadays, there are things that are so real that we cannot distinguish them with the naked eye. I myself, a long-time professional, also have content that is very difficult to distinguish, let alone for ordinary people," Mr. Hung shared.
The worrying thing is that deepfake does not need to reach perfection to become a fraud tool.
According to Mr. Hung, deepfake does not need to be perfect and can still deceive the victim if the bad guy creates an emergency situation, causing the recipient to lose their temper and make hasty decisions (Photo: Tien Dat).
According to Mr. Hung, bad guys often put victims in urgent situations: relatives encounter problems, need to transfer money immediately or have to make a decision in a very short time.
When emotions are heightened, it is difficult for the recipient to be calm enough to observe every small abnormality on the face.
Therefore, instead of just learning how to "scan" a fake video, Mr. Hung believes that users need to form a more important reflex: slow down and verify information using another channel.
As a face on a screen can be made increasingly more realistic by AI, “seeing someone familiar” will no longer mean that person's identity has been authenticated.
Photo: Tien Dat