What is Ring and what happened?

Ring is an American company that produces several products, including smart door bells. They have a built-in camera that can broadcast video to the user’s smartphone. In order to improve the capabilities of their devices, the Ring team used these videos to train neural networks: cameras could, for example, distinguish a person from a dog, a passerby from a car in motion.

This is what the Ukrainian team Ring Ukraine started doing. In 2016, the R&D center of the company was opened in Kyiv, one of the tasks of which is the neural network training.

According to the American media, Ring Ukraine employees had access to all user videos, since they were transmitted in an unencrypted form, and could use them for other purposes: for example, download them.

Where did the data come from?

Sources “The Information” and “The Intercept” differ. According to Vector which was the first to draw attention to the publication of “The Information” in Ukraine, the publication announced a survey of 24 current and former Ring employees. It is not reported how many people exactly responded to the requests of the publication and whether all 24 people confirmed the statements of the publication. Judging from the article of “The Intercept”, the source of the information were two former employees of the company.

According to sources of AIN.UA, one of those employees who gave the comment to “The Information”, was fired from the company after six months of working there. He held a fairly high position, however, after several cases of fraud, the company said farewell to him. As a result, he turned to the press, remaining negatively disposed towards the company. Mind has a similar version, according to the sources of which, a former employee of Ring Ukraine blackmailed the company and did not receive his due, he referred to the press.

So, was the access to the video or not?

Ring noted that all data is actually encrypted and there is no free access to the video. The company does not give any comments to the rest of the questions. The Ukrainian office of Ring Ukraine refers to the statement of the American press service.

The editorial of AIN.UA was able to talk with two former employees of Ring Ukraine and find out what is the algorithm of work with videos.

Kostiantyn Bokhan is a former research director of Ring. He worked there from April to November 2017. According to him, videos were stored on Amazon servers.

“Every American Ring client has access to his/her video through a web interface. And there he/she can indicate that he/she could share this video. Then this video was transferred to another folder, again, on Amazon servers. From there, the materials were taken for neuron training. The entire backend Ring is hosted on Amazon servers. Accordingly, all the work being done in Ukraine is also being carried out on these servers.”

The job of the operators who trained the neural network was to mark the objects in the video: the face, the car, the moving object, and so on. Operators in the Ukrainian office worked with a special video system that had access to that folder on the servers. This system transmitted video, which is allowed for markup, to the operator on the screen. The operator indicated what he/she had seen, and all this data was sent to the Amazon server again.

Image: The Intercept.

Accordingly, customer materials were not physically stored in the Ring office. These are huge amounts of data; the company would have to buy a data center here to store and process it. So, Amazon is responsible for protecting this data. Ring bought Amazon services for both data storage and processing. All processing procedures were launched in the United States on the servers of the latter.

To access these videos, the device of the staff was specifically configured. It is required by the office to include encryption of all drives.

Another employee of Ring Ukraine, who wished to stay anonymous in his comment, had been working in the company as a developer from June 2017 to August 2018. He also denies the allegations made in the American press.

According to him, the operators who viewed the videos could not download or send them to anyone. A specific computer could get only a specific list of videos available for processing and besides direct tasks, it could not do anything with this video.

“We made an application for operators. This is how it works: in order to check how well the AI ​​determines an action, it is given the same list of frames for analysis, as the operator receives. The person knows where and what the action is and in order to check how correctly the AI ​​learns, he is given a list without markup, and then it is checked. The comparison is a sample and we understand how accurate or not AI is, and what needs to be done to make it work more accurately.

The operator, who worked through the application, does not know where the materials came from, from whom they are received, and so on. Moreover, he cannot save him. He has one common task: to mark a person on it,” added the commentator.

The editors of AIN.UA did not manage to obtain information whether all these procedures were followed in the Ring and Ring Ukraine companies from the very beginning.

Is it normal to teach AI for two years?

During the preparation of the article, the editors of AIN.UA faced another question from users: how did it happen that the Ring team has been teaching the neural network for two years now? Does it take so long? Is this an AI training, or is the company just saving money, or does it simply have no technology?

According to the AIN.UA edition, several people who are engaged in the development of their own projects with neural networks, say that one can teach the network as long as possible. And the better you manage to train it, the better it will work. Moreover, it is really not an easy task.

One of the sources of the publication, who wished to remain anonymous, said that Ring, most likely, had a very difficult task in training the network. For example, if a tree fell in front of a house, you need to teach the network that it was not a person who passed, but a tree fell, same with a car driving by, or a cat running around. And if it is not a cat, then it’s a dog, and you have to teach it that each animal must be distinguished from the other.

“For people who do not understand the process of training a neural network, it is difficult to understand what this process actually is. After all, any factor influences its performance: how the ring was hung up, at what angle, whether it is raining, snowing or the sun is shining. And you need to understand that there is a man walking, and there is a man with a dog, and there is a man near whom a bird flew. Is this a man in the hood or does he have a hairstyle like that? All this takes time and you need to understand that the more data you give to the neural network, the better it will work,” said the source AIN.UA.

Oleksandr Arapov, VP Engineering of 3DLOOK company, in the commentary to AIN.UA confirmed that the training of a neural network on video streams in any case, takes a lot of time.

“There is a manual factor in data preparation: the more data you have, the more manual labor or process automation is required, and automation is development, which is also equal to time. After data preparation, there is a training stage, and everything depends on resources there. Judging by the fact that Amazon bought Ring, there are no problems with resources. However, since they work with videos, it takes time to train the neural network on video streams regardless of the purpose. Plus, if there are different goals and different data, there is always a need to educate, retrain, optimize the network, transfer it to the device. Here you have 2 years,” said Arapov.

At the end, Oleksandr added that it is training of the neural network on a huge amount of data followed by its implementation in products that allows Ring to differ from competitors who could not give their neural networks a large amount of information.

The user agreements reviewed by AIN.UA editors indicate that the owners of Ring door bells gave the right to use their video.

The article was prepared by the editors of AIN.UA: Dmytro Demchenko, Ilya Kabachynskyi,  Ilya Boshnyakov