Shulou(Shulou.com)06/02 Report--

The anti-deepfake camp has new achievements today. Shang Tang entered the bureau and presented the largest detection data set so far:

It contains 60000 videos, totaling 17.6 million frames, 10 times more than the existing dataset.

Deepfake evolved version after version, the effect became more and more realistic, but the threshold became lower and lower.

△ Schwarzenegger, Bill Harder without trace face change

Not only the celebrities, but even ordinary people couldn't help but tremble.

Seeing is not believing. Is there nothing that can stop AI from changing faces?

In fact, when the devil was one foot high, the Tao did not stop cultivating. And he will return to his ways.

Now, Shang Tang has teamed up with researchers at Nanyang Technological University in Singapore to launch the largest deepfake detection dataset to date, Deepforensics-1.0.

Moreover, it is closer to the reality scene, more diverse and challenging.

Data, code and pre-trained models are on the way to open source.

DeeperForensics-1.0

Of the 60000 videos in DeepperForensics-1.0, 50000 were original videos collected by the research team, and the remaining 10000 were "fake videos" they created.

The creation of the dataset was divided into three steps.

The first step is data collection.

By calling the original face in the real video the target face and the replaced face the source face, the researchers found that the source face played a more critical role than the target face in building a high-quality dataset.

The richer the expression, pose and lighting conditions of the source face, the higher the reliability of face exchange.

So the researchers hired 100 actors to participate in the recording of face videos. They came from 26 different countries, 53 men and 47 women, ranging in age from 20 to 45 years old, with a 1:1:1 ratio of four skin colors (white, black, yellow, brown).

These videos were recorded at a resolution of 1920×1080. During filming, actors were asked to display various expressions: neutral, angry, happy, sad, surprised, disdainful, disgusted, fearful, etc.

The angle of the face facing the camera varies from-90° to 90°. Nine different lighting effects are also set.

The second step is to fake it.

Know yourself and know your enemy.

To generate more realistic fake videos, the researchers proposed a new framework for face swapping: DeepFake Variational Auto-Encoder (DF-VAE).

DF-VAE consists of three modules: structure extraction module, decoupling module and fusion module.

In training, the source face and target face are reconstructed by extracting markers and constructing unpaired samples as conditions.

After reconstruction, optical flow differences are minimized to improve temporal continuity.

The MAdalN module is responsible for blending the reconstructed faces with the original background.

The third step was to further increase the difficulty by adding disturbance to simulate the video in the real scene.

Specifically, seven distortions are added to the video: color saturation change, local image block distortion, color contrast change, Gaussian blur, Gaussian white noise in color components, JPEG compression and video compression ratio change.

To assess the quality of Deepforensics-1.0, the researchers asked 100 computer vision experts to rate it.

Based on feedback, experts believe that Deepforensics-1.0 is more realistic than popular Deepfake detection datasets such as FaceForensics+ and Celeb-DF.

Blocking Deepfake

The fake video became more and more real, causing widespread concern.

The AI has already begun to act.

Earlier, Facebook threw tens of millions of dollars to hold a face-changing video detection challenge.

UC Berkeley EECS Professor Hany Farid commented:

In order to move from the information age to the knowledge age, we must do a better job of identifying the true and punishing the false, and educating the next generation to be better digital citizens. This will require comprehensive investment and a concerted effort by industry, academia, and NGOs to research, develop, and implement technologies that can quickly and accurately identify authenticity.

Truepic, an American start-up focused on cracking down on AI fake photos and videos, raised $8 million in July 2019.

At home, the Regulations on the Management of Network Audio-and video-based messages issued at the end of November 2019 can be regarded as a targeted control over AI fake videos.

This regulation came into effect on January 1.

the portal

Project Address:

https://liming-jiang.com/projects/DrF1/DrF1.html

Address:

https://arxiv.org/abs/2001.03024

VB reports:

https://venturebeat.com/2020/01/15/sensetime-face-forgery-research-deepfakes/

- End-

https://www.toutiao.com/i6782403843717071368/

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.