The Center for Data Innovation spoke with Marian Gläser, chief executive officer and co-founder of Brighter AI, a Berlin-based startup which developed a deep learning software to use publicly recorded camera data for analytics and AI. Gläser discussed how Brighter AI’s solution empowers companies to use video data for various sectoral innovations, while being compliant with privacy regulations.
Eline Chivot: What is your background as a tech entrepreneur, and what inspired you to start Brighter AI?
Marian Gläser: I’m a computer scientist by training and I first got into tech entrepreneurship while working for a startup in the city committee. Back then we were developing 3D scanning models. It was the first time I noticed that deep tech has a huge impact on how industries develop, especially in computer vision applications. That inspired me quite early on. From there, I went into further engagement in the startup scene. I deep dived into IT management and consulting to get an understanding of the basic tools regarding how companies and IT operate and work side by side in a startup.
I was scouted by Hella, a large automotive company. Back in 2016 they were seeing great opportunities in the emerging trends, such as electrification and autonomous driving, transforming the mobility world. Along with other automotive companies, Hella created incubators and accelerators to respond to those trends. I worked there as an “intrapreneur”—an internal entrepreneur, to investigate disruptive innovation within the automotive field. Anything that could potentially disrupt Hella with a new business model or technology would be exciting to investigate.
At some point, we started investigating neural networks and how they could disrupt existing business models in the sensor field. As a result, we developed a neural network that could turn a night-vision camera image into a day-light image. Hella is very strong in headlight manufacturing and technologies. The transition from night-vision camera into day-light image is extremely useful for driver assistance systems. Then you might not need headlights as much anymore, meaning this has considerable disruptive potential. It was an exciting project, which inspired me to start Brighter AI—and which inspired the name of our startup.
Chivot: How exactly does Deep Natural Anonymization Technology (DNAT) work? Is it more resource-intensive than traditional pixelization or image annotation?
Gläser: DNAT is a solution protecting personally identifiable information, such as faces and license plates, in public. When using traditional anonymization methods, such as pixelation or blur, data loses the ability for further processing and applying analytics or machine learning. When a vehicle recording video training data notices a person crossing the street, specific information such as line of sight is needed to determine whether the person is aware and saw a vehicle or not. Furthermore, extracted data such as age and gender enable modeling predictive behavior.
The moment you blur or pixelate a face using traditional anonymization methods, this information is lost. Basically, none of the above would work for industries such as automotive, where advanced analytics is crucial. We solved this by generating synthetic overlays for faces and license plates which protect the identities, as you cannot reverse-engineer them to return to the original image. At the same time, all the attributes are kept. We are able to extract 25,000 data points of facial expression and pixel-perfect line of sight for further analysis. Generic data such as age and gender is also preserved, making the person fully understandable for neural networks while keeping the identity protected.
Regarding resources, we do have a very complex technology. However, we can break it down in the software footprint to minimum. In consequence, it does not use that much overhead, compared to traditional anonymization methods. Added value that we bring is the anonymization. With our solution companies can collect data in public and make us of it for new business models or cost-reduction. So usually, the overhead of additional calculation is essentially negligible when determining the value.
We have internal processes to ensure that natural anonymization is not reversible. The neural networks themselves are trained with a randomized factor without a connection to the original image. So, it is technically impossible to go back to the original face. We also ensure this with third-party checks, which means that we have a couple of state-of-the art facial recognition software that we try against our anonymization, to test every release that we do. We release regular upgrades that have a higher performance and a higher coverage of protection, and before these are released, we heavily check them against third-party facial recognition.
Chivot: Can you give examples of how DNAT can transform the automotive and retail industries? And are there any other areas or industries where DNAT can be applied to?
Gläser: DNAT could change industry’s perspective on privacy, the way the GDPR did. When we started to develop our natural anonymization model, we noticed that within the automotive industry, a lot of companies felt that the GDPR was limiting their innovation, processes and development. We felt that there were technologies, such as neural networks, able to resolve those arising privacy barriers. DNAT revolutionizes the automotive industry to the point that companies can collect video data on the street without having to go through the legal processes concerning data protection. Removing this part of their efforts—investing resources to comply—enables companies to collect more data faster, share this data with their partners more easily, and therefore accelerate the deployment of innovative projects.
DNAT could be applied to a lot of industries. We started in the automotive field with testing and mapping vehicles. Then we noticed that not only testing vehicles in their first development phase are collecting data. Nowadays also in-series production vehicles do that—they are collecting data and streaming it back to the server to the tier-one access management. The most famous case is Tesla, which publicly announced that with eight cameras, they are streaming video snippets from vehicles back to Tesla every time the driver would disengage the autopilot. Other companies in Germany and Europe have exactly the same approach of continuously collecting data from feeds.
DNAT holds many promises for smart cities. One of the clients we onboarded last year is Deutsche Bahn. With around 40,000 cameras they were not able to use any of the data that was being recorded, except for security purposes. That is quite astonishing if you think about it, because these thousands of cameras collect data which could be used for smart cities. For example, what type of passengers are travelling on a train and walking on the platform? Are there ways to improve train or bus stations based on people’s demographics? We can also use this data to better understand the exact behavior around any type of tracks in emergency cases such as evacuation. Much of public transport revenue depends on the shops and services integrated into the train stations, but they do not get to gain an understanding regarding who is actually walking into those stores and who is attracted to a particular space more than to another. Such an understanding of passengers would allow these stations to improve their journey and experience through their services. When we first integrated DNAT into a DB camera and started testing, one employee said that it was a historic moment for the Deutsche Bahn to have such a technology in their stack.
Another use case in retail are spaces such as Amazon Go concept stores, that heavily rely on camera usage. As cameras are one of the richest sensors in terms of data potential and, at the same time, are rather cheap to require, there’s a great potential for us. Public areas where there is some kind of added value or cost reduction opportunities being created with camera usage, are where you start grappling with privacy. That is something that plays into our strategic advantage.
Except from Europe, we see a similar approach towards privacy, especially regarding big data collection, in the United States. Companies are aware that they have a social responsibility regarding this data. Beyond regulatory obligations, they have an intrinsic motivation to think about how to protect that data, for instance against data breaches. So, while we are working in automotive and public transport, we also seek to expand into other regions around the world. We just got our first clients in Japan and in the United States, so we are very happy to be present at the right time, with technology that comes from Europe.
Chivot: Can you give an example of how cities have adopted DNAT, and what have been the results?
Gläser: We started with a first pilot for Berlin’s public transport, and we are now going into the next steps of rolling it out. This will open up the potential for the city to use cameras. It started with very simple analytics and now the city will use them for more advanced data points, such as the counting of passengers. They plan to understand the demographics in order to transform part of the infrastructure based on more data points. This will have a huge impact in terms of what public transport users in Berlin will be able to experience in a couple of years, without any concerns regarding their privacy.
Chivot: What does the future look like for Brighter AI and the anonymization industry in general?
Gläser: Further improvements may result in making more elements of a human body or image personally identifiable. We started with license plates and later on facial replacements, as facial recognition has improved over the past five years to above 99 percent in accuracy. In other words, if your or anyone else’s photo can be matched to a database with facial recognition, you can track this person with high certainty. There is promising research developing a system identifying a person by the way he or she walks. The recognition rate isn’t as strong as it is for facial recognition, but it is certainly an area where we see improvements. We are continuously researching more elements in image data and automatic recognition which we can both identify and anonymize. Very likely, something such as understanding the ways in which people walk and identifying a particular person through that will become just as straightforward as using traditional optical identifiers such as tattoos or stickers on cars can be.
Europe has a great potential to become a market leader in privacy tech. There are many privacy-tech companies emerging with solutions for the anonymization of data or synthesized data and systems that require less personal information. We see in the visual space that more companies are working on anonymization, and we see anonymizing web traffic and personal details. These technologies will be crucial in the future, as many international companies are adopting privacy-oriented solutions.