Author Archives: admin

They trick companies by simulating CEO voices with AI

Scammers mimic the CEO’s voice with deep-fake. Almost perfect scam.

A British energy company has been the victim of an unprecedented and complex variant of the “Presidential Scam.” Keep in mind that this scam involves impersonating the head of the company and prompting an emergency transfer to a mob-controlled bank account. In this case, the scammer relied on artificial intelligence systems to perform a nearly perfect simulation of the voice of the German parent company’s CEO, which was disclosed by the Wall Street Journal and relayed by Gizmodo. In the field of artificial intelligence, this simulation is also called “deep forgery”.

The victim was questioned by the insurance company Euler Hermes, who explained that the artificial voice mimicked not only his boss’s German accent, but his voice. Needless to say, the poor man nodded immediately and ordered the transfer of 220,000 euros to a bank account in Hungary within an hour. The giants were then transferred to Mexico and then to other unknown destinations. Euler Hermes said this was the first time he had encountered an artificial intelligence-based scam. May not be the last.

Researchers rally to twist Deepfakes

Facebook has just officially launched its “Deepfake Detection Challenge” to advance research in the field … and eliminate this new plague in its social networks. The hunting season is open.

Not all technological advances are good, and the audio-visual forgery technology “Deepfake” based on neural networks is particularly toxic. It can change the heroines of pornographic movies, modify the speech of politicians or celebrities, and even create scams. Welcome to the world of lies and camouflage.

No wonder Facebook has already faced a wave of trolls and intoxication, and is very interested in blocking such content. The social network took advantage of the NeurIPS conference in Vancouver from December 8th to 15th to officially launch its Deepfake Detection Challenge.

$ 10 million reward
The idea is to give AI researchers the body of more than 100,000 “Deepfake” videos and encourage them to develop detection technologies. Google has created a similar corpus, but the scope is much narrower. In addition, Facebook went further, offering $ 10 million in scholarships and rewards to the best teams … as long as they released their code in open source.

* More than 150 teams have registered. They must submit the solution by March 2020 and will evaluate it with unpublished data sets. “This is a huge challenge for Facebook,” Antoine Bordes, co-director of Facebook AI Research, told a press conference in Paris. “We are already developing internal detection methods, but we cannot do it alone everything. That’s why we try to openly unite communities around this issue. ”

To create this dataset, Facebook hired actors and filmed them in various situations. The company then produced the deepfakes for these records using the best technology. In the example provided by Facebook above, can you determine which was the original video? (Tip: look closely at the eyes and their outlines *).

How does Deepfake work?

Deepfake was created based on two types of cascaded neural networks: “autoencoders” and “generative adversarial networks” (or GANs, generative adversarial networks). An autoencoder is an algorithm consisting of an “encoder” and a “decoder.” The first was able to reduce the image to a minimalist mathematical representation called a “latent image.” The second is able to reconstruct an image based on this representation. To replace Clara Morgane’s head with Scarlett Johansson’s, all you have to do is combine the first encoder with the second decoder. And voila.

How does Deepfake work

Autoencoders can also be of variable type. By simply changing features based on the probability curve, they can generate similar images instead of simply recreating well-known images.

The disadvantage of the encoding-decoding process is that information is always lost. In the case of change generation, blurred areas also appear. Therefore, the researchers came up with the idea of ​​combining an autoencoder with a GAN whose task is to deliver only images of sufficient quality.

In fact, the generated adversarial network consists of two competing neural networks. The first generates images, and the second must determine whether the images are true or false based on his training. The more complicated the discriminator’s analysis, the more “generated” the generated image will be. For Deepfake generators, the decoder of the autoencoder can simply be replaced by a GAN generator.

To detect Deepfake, researchers will once again rely on deep neural networks, which they will configure and train to be able to identify the truth from the errors. So interested in having a big data set like Facebook has to offer. Deepfake detection is an emerging field, although there are already a dozen different methods. Some researchers focus on blinking, others focus on temporal inconsistencies between images, and others focus on their texture and resolution.

Towards an endless race?

This study will obviously lead to the shallot race. As the detector improves, the quality of the generator will improve. Is it possible to win this search match? Facebook grew there. “Look what’s happening with spam. The filters are running and the problem has been resolved. In Deepfake, the situation will be the same. The more sophisticated the detection technology, the higher the cost of undetectable forgery. You have to lease the computing cluster It takes genius to make great Deepfake videos, but people will stop doing it. “Antoine Bordes said.

In other words, in the Deepfake competition, the winner will eventually be the one with the most computing power. It’s hard to beat Facebook in this regard.

5 Robots to learn Coding

Learning programming with robots is more fun. This is our choice of five models for different age groups.

Providing a robot for programming is very expensive. Count 100 euros first. And don’t think this is reserved for children. This year, new products have appeared in the entire family.

Ozobot: The easiest

This robot allows you to learn coding from the age of six or less, as it is not absolutely necessary to know how to read and use it. The starter pack includes a robot with accessories including code sheets, markers, sticker boards and activities. The child just needs to draw a color code or paste a sticker to send the information to the robot that will be placed above the sheets. Equipped with sensors that can recognize lines and colors. Then it will go forward, back, turn, slow down or accelerate. Simple and efficient.

Price: 80 Euro

Thymio 2: most educational

Thymio is already in use at the school and it was developed in collaboration with the Federal Institute of Technology of Lausanne and the Lausanne Academy of Arts. Equipped with a large number of sensors, actuators and LEDs, it can be programmed through Aseba Studio. It is open source and can also be used with Lego and controlled by Scratch software. Thymio already offers many educational activities online, such as calculations and geometry exercises.

Price 139 Euro

Blink: the cutest

He is the latest member on the list. After successfully completing the crowdfunding campaign on Ulule, he came to this Christmas with a label made in France and sponsorship from Ecole Polytechnique. Suitable for children from 5 to 12 years. Since it is provided separately, all sensors must be assembled and read first. Then use it from an application that allows you to write games using different modes depending on the complexity. So Winky can be turned into an alarm clock. Once the program is created, the child can interact directly with the robot without going through the application. Special attention was paid to the two ears with activation function, low noise and accessories for changing the appearance.

Price: 200 Euro

RoboMaster S1: suitable for older children

Don’t make any mistakes. This so-called “educational” robot can be used by adolescents, but is best assisted by older robots. It is also the ideal toy for adults. Because you must at least master the Scratch programming language now taught at French universities. More ambitious people can also write code in Python. The RoboMaster Lab programming interface is accessible from any screen. The robot looks like a tank, has four wheels and a turret, and can face other players. It is equipped with sensors, capable of drawing environmental maps, identifying and tracking individuals, moving in all directions and even projecting rubber balls into ammunition. Obviously, all of this comes at a price. The RoboMaster S1 actually costs 550 Euros, which does not include accessories sold as a kit, such as “Play More”, which costs 170 Euros, including the controller, battery and projectiles.

Price: 550 Euros

Sphero RVR: Suitable for most DIY enthusiasts

The latest member of the Sphero brand is aimed at students, scientists and anyone who wants to tinker. It is not distributed directly in France, but can be purchased in US dollars on the official website or It has been a huge success on crowdfunding website Kickstarter. This very simple rover is controlled by the Sphero Edu app and can be programmed from iOS or Android. Its interest lies in its excellent modularity, as third-party components and accessories can be added by connecting a Rasperry Pi or Arduino to the upper platform of the machine.

Price 227 Euro

Millions of Android smartphones infected with adware hidden in hundreds of applications

More than 100 apps currently available on the Play Store are hiding adware already installed on millions of devices.

White Ops, a company specializing in cybersecurity, has published a study showing that 100 applications available on the Google Play store and downloaded by more than 4.6 million devices contain adware that broadcasts ads (Even if the app in question is closed).

Millions of infected smartphones

The new malware features silent advertisements on the user’s own device and pop-up advertisements on the user’s device. All affected applications share a code library called “Soraka” (from which the malware got its name) and a variant called “Sogo”.

Among the malicious applications discovered by White Ops is Best Fortune Explorer, an application released last September. Passed unhindered without being affected by antivirus scanning and already accounted for more than 170,000 downloads.

Adware overlooked in the Play Store

The mode of operation of Soraka is quite sophisticated. The code found in these infected apps indicates that a filtering system has been installed to determine if certain conditions have been met to display unwanted ads on the user’s device.

The advertisement page is not displayed immediately after installing the fake app. Soraka waits a bit before triggering to avoid detection by various antivirus scanning tools. When the user unlocks the screen, the ad still appears, but there are no open applications, which confuses the problem and prevents the user from identifying the underlying application. 1 2 3 4 5 6 7 8 9 11 b c d g h I k l m n o p q r s t r so x

In addition to a detailed description of how this adware works, White Ops publishes on its website a list of 104 Android applications (unfortunately little known) affected by Soraka.


Source :
White Ops

What are the ten most downloaded applications in the past decade?

Between 2010 and 2019, the ten most downloaded apps were primarily communications related, while Netflix and Tinder were the two most used apps in the world. Find the top ten.

AppAnnie, a research company specializing in mobile behavior, first released the world rankings of the top ten apps downloaded in ten years.

Facebook group on the podium

One of the salient features of this study: 7 of the 10 apps are communications apps, of which the first three are Facebook, Messenger and WhatsApp. Instagram ranks fourth. Therefore, the four most downloaded apps in the world since 2010 belong to Mark Zuckerberg. Witness its amazing achievements among young people: the recent TikTok app has risen directly to seventh place, second only to the long-established Skype.


The ten most commonly used applications in the world are …
At the same time, the company categorized the ten applications where users spend the most time. Not surprisingly, video or music streaming applications top the list. Netflix is the most time-consuming application, followed by Pandora Music and Tencent Video, ranking third. This also illustrates our mobile habits: Tinder is the second application that users spend the most time on.

+ 15% of app usage per year

The mobile support market is booming. According to AppAnnie, downloads grow by 5% per year, and time spent on mobile devices grows by 15%. Basic trends cannot change.