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

Introduction: an invitation letter from Facebook

Lei Feng AI developer's note: just last month, Facebook launched the Deepfakes testing Challenge, offering a reward of $10 million to fight fake AI face changes.

Now, with a new move in the challenge, Facebook has found like-minded allies such as AWS and Microsoft to release the initial data set of the challenge and launch invitations to AI bosses around the world. Facebook released an article to describe this development in detail, and the AI developers of Lei Feng.com compiled the relevant content as follows.

Progress of Deepfakes Challenge

It is very important for the whole industry and the whole society to develop the technology to detect false video effectively. Similarly, the development of these tools also requires the support and cooperation of experts in the field of artificial intelligence. To accelerate this important work, we have now released the first subset of more than 100000 videos created specifically for the Deepfakes Detection Challenge (DFDC).

We released the DFDC program last month to promote the development of new tools to better detect where artificial intelligence is used to change videos or images, causing misunderstandings among viewers. Now, we are also pleased to announce that the DFDC project has been supported and joined by Amazon Web Services (amazon web services,aws) and several other leading academic researchers.

Initial dataset release to obtain feedback from the research community

The DFDC dataset will contain a variety of new, high-quality videos produced specifically for the study of the Deepfake Challenge. These videos are recorded using only paid actors who have reached an agreement with us and helped us create data sets to avoid restrictions that could hinder the work of researchers. Information on data sets and rankings, as well as bonuses and awards funded by Facebook, is available on the DFDC website.

To ensure the quality of data sets and challenge parameters, we are now sharing the first 5000 DFDC videos with researchers in this field. Starting from October 27, we will collect feedback and hold targeted technical working sessions at the International computer Vision Conference in Seoul. The full dataset release and DFDC launch will be launched in (neurips) in December this year.

Currently, one of the components of the DFDC data set is made up of tens of thousands of videos containing faces, with tags describing whether they were generated using facial manipulation technology. All videos in the dataset are created with the consent of paid participants, and the dataset will be made available to the community free of charge for developing, testing, and analyzing techniques for detecting videos with manipulated faces.

DFDC website:

Https://deepfakedetectionchallenge.ai/

Dataset address:

Https://deepfakedetectionchallenge.ai/dataset

New members of DFDC

We recognize that communities need to work together to build tools to deal with Deepfake more effectively. Among them, more industry partners, as well as academia, media and social organizations will be the key to achieve this goal.

AWS also joined the work as a technology partner and as a member of the New steering Committee on artificial Intelligence and Media Integrity (Partnership on AI "s new Steering Committee on AI and Media Integrity). The committee will oversee this challenge and will consist of a broad coalition of cross-sectoral organizations, including Facebook, Witness, Microsoft, the New York Times and other social, technical, media and academic organizations.

AWS has not only contributed up to a $1 million AWS credit line to support this challenge, but will also have technical support and guidance from experts from Amazon ML Solutions Labs. After the challenge, AWS will also work with the participating team and provide the opportunity to host their models in the AWS model market based on the contestants' personal wishes.

AWS blog:

Https://aws.amazon.com/cn/blogs/machine-learning/aws-supports-the-deepfake-detection-challenge-with-competition-data-and-aws-credits/

The voice of more scholars

Professor Pietro Perona

"as a long-term researcher in the field of artificial intelligence, I believe that artificial intelligence can and should be a useful force in the world! Pietro Perona, a professor of electrical engineering and computer and nervous systems at the California Institute of Technology, is both a researcher at AWS and will serve as a technical adviser to DFDC.

He told us: "to combat the use of Deepfake as a tool to deceive the public and threaten society and democracy, we must face it bravely." That's why I'm keen to help organize the Deepfake Inspection Challenge and work with Facebook, the Partnership for AI and others to create a community that helps mobilize machine learning and computer vision to compete. It is precisely because competition unites us, helps us sort out the problems to be solved, identifies the steps we can take, and sets a standard for subsequent progress to help the world maintain a bright line between reality and fiction. "

Professor Laura Leal Taix é

Professor Laura Leal Taix é, head of the dynamic Vision and Learning Group at the Technical University of Munich, also joined us as an academic adviser and collaborator as we developed datasets for the wider community. She and Cristian Canton Ferrer of Facebook artificial Intelligence will host the DFDC conference at the International computer Vision Conference ICCV.

"the technology for generating real images is improving at an alarming rate. While this is incredibly exciting from a technical point of view, there are obvious social concerns, and fake media videos are clearly one of them. This may not seem important for a generation that is not yet fully digital and dependent on print media, but future generations are more likely to consume news only in digital format. And it is completely vulnerable to the proliferation of fake news and target media. "

Leal-Taix é said, "therefore, it will be a great challenge for society to educate this new generation of media consumption and raise people's correct understanding of the false news that already exists around us. From our point of view, we must work harder to create new technologies to combat digital forgery, which can only be achieved by the combination of industry and academia. I think DFDC is an outstanding effort in this direction! I hope this will inspire thousands of wise people to solve this problem. "

Professor Luisa Verdoliva

Professor Luisa Verdoliva of the Department of Industrial Engineering Federico II of the University of Naples will also join DFDC as an academic adviser. In addition, Professors Leal-Taix é and Verdoliva will work with other leading academic researchers to help build this challenge.

We believe that this open, community-based effort will accelerate the development of new open source tools to prevent people from using artificial intelligence to manipulate videos to deceive others.

Frequently asked questions (FAQ)

What is the goal of the Deepfake testing challenge?

AI technology related to Deepfake and other tampered media is developing rapidly, which makes it more and more difficult to detect Deepfake, and sometimes even human evaluators can not reliably tell the difference between the two. The Deepfake testing Challenge aims to create new ways to detect fake videos, prevent the public from being misled by the media, and stimulate the rapid development of the field by inviting participants to compete.

When does the challenge begin?

With the release of the extended dataset, the Deepfake detection challenge will be launched in December 2019.

When is the deadline for submission?

This challenge will continue until the end of March 2020.

How does the challenge proceed?

Participants can download the dataset used to train the model and then submit the code to the black box environment for testing. We will begin to submit the challenge later this year, when guidelines and dataset licenses will be provided.

How do I create a training dataset?

We are building a new training data set specifically for this challenge. To create this dataset, we are working with a third-party vendor who employs a different group of people who have agreed to participate in creating the dataset for this challenge. Then we use a variety of artificial intelligence technologies to create tampered videos based on a subset of these unmodified videos.

Who can take part in the challenge?

The challenge will be global and participants will need to agree to our dataset license before participating in the challenge.

Do you use user data from social media or video platforms in your dataset?

The training data set does not include user data for social or video platforms. We are building a new dataset for this challenge.

How to judge the challenge / how to choose the winner?

We will provide a testing mechanism that enables teams to rate the effectiveness of their models based on one or more black-box test sets provided by our founding partners.

What are the rights of challenge participants to the technologies they have created for the challenge?

Participants will reserve the right to train models on the training data set. Facebook and its subcontractors will be given the right to use these models to manage the challenge.

How to guard against opponents who try to access code and data?

We will restrict access to the training data set so that only challenged researchers can access it. Each participant needs to agree on terms of use for how to use, store, and process data, and there are strict restrictions on sharing data.

For more details, visit the DFDC website:

Https://deepfakedetectionchallenge.ai/

Https://www.leiphone.com/news/201910/ikJtvoIeZ62Te4kD.html

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.