Rise of AI Summit 2020 · Berlin · Society & Strategy Day | Future Space

Combining Privacy and AI

Thomas Strottner · Head of Business Development, brighter AI Technologies GmbH

At the Rise of AI Summit 2020, Thomas Strottner, Head of Business Development at brighter AI Technologies, filling in for the company's CEO, explains how his company reconciles data privacy with AI development on camera data. He describes cameras as the richest sensor for AI use cases such as Amazon Go stores, smart city people counting and self-driving cars, and notes facial recognition systems can now track up to 20,000 key points on a face, which brings both useful applications and risks of abuse, citing facial recognition used to track Black Lives Matter protesters and for social scoring in authoritarian regimes. He surveys tightening regulation, including GDPR and California's new Consumer Privacy Act, and describes a perceived "AI privacy dilemma" between accurate models and privacy protection, which he argues is solvable through approaches like federated learning, privacy by default and privacy by design. He explains that pixelating faces and license plates degrades AI training data and introduces bias, and presents brighter AI's alternative: generating synthetic faces and license plates that preserve realism while, in tests against Microsoft Azure's facial recognition, reducing identity match to about 16%.

Key takeaways

  • Strottner says modern facial recognition can detect up to 20,000 key points on a face, enabling both useful applications and abuses like protester tracking and social scoring.
  • He presents federated learning, privacy by default and privacy by design as concrete ways to ease the perceived trade-off between accurate AI models and data privacy.
  • He argues pixelating faces and license plates for anonymization introduces artifacts and bias into AI training data, such as models expecting pixelated plates in real driving.
  • Brighter AI generates synthetic faces and license plates that, in tests against Microsoft Azure's facial recognition, matched the original person only about 16% of the time.
  • He frames California's GDPR-like Consumer Privacy Act and Europe's regulatory head start as a business opportunity for European privacy-centric AI solutions.

Detailed notes

Cameras as the richest AI sensor

Strottner introduces brighter AI, which recently won the Berlin Deep Tech Award, and frames camera data as the richest sensor for AI use cases: Amazon Go's automated checkout stores, smart-city applications like people counting, and self-driving cars, where he notes Tesla emphasizes cameras as the most important sensor. He argues AI development needs both data quantity and data quality, and that privacy is often seen as standing in the way of both.

Facial recognition's promise and abuse

He points to facial recognition as one of AI's notable achievements, such as unlocking a phone with a face, enabled by modern neural networks that can now detect up to 20,000 key points on a face. He balances this against real risks of abuse, citing facial recognition used to track individuals at Black Lives Matter protests and its use by authoritarian regimes for social scoring, alongside growing individual and societal pushback on privacy.

A tightening regulatory map

Strottner presents a map of global data protection regulation strength, noting the US, Canada, European countries under GDPR, and even China are marked as having heavy regulation, which restricts how companies such as automakers can collect camera data in public. He highlights California's new Consumer Privacy Act, closely modeled on GDPR, as an opportunity: because California is one of the largest economies and home to major tech companies, he sees a chance to export European privacy-centric solutions, noting San Francisco has banned facial recognition technology. He references a study brighter AI conducted with the German AI association on video data regulation across countries, stressing that internationally operating companies must track diverging national rules.

The AI privacy dilemma and how to address it

He describes a commonly perceived dilemma between powerful, accurate AI models built on clean data and privacy protection achieved by degrading data, such as blurring faces. He presents several approaches to ease this: decentralized processing such as federated learning, where models train across distributed devices without exchanging raw data, as used by Apple; privacy by default, which sets the most privacy-protective settings as standard, as GDPR requires and as seen in cookie banners; privacy by design, building protection into the technology itself, for example through encryption, pseudonymization or full anonymization; and altering the data itself through synthetic data or data augmentation, the area where brighter AI operates.

Synthetic data and brighter AI's synthetic face solution

For self-driving cars, he discusses fully synthetic, generated-from-scratch data and digital twins as ways to avoid privacy issues entirely, though he notes synthetic scenarios can never capture the full creativity of real-world edge cases, and that data augmentation, adding artifacts like fog to existing clean images, can help generate more corner cases from privacy-compliant data but can never fully reflect reality and can introduce bias. He argues the common alternative, pixelating faces and license plates, introduces artifacts, such as a segmentation model detecting a "phantom person" or a bird where pixelation appears, and can bias models, for example training a network to expect pixelated license plates that do not exist in real driving. Brighter AI instead generates synthetic faces and license plates that look natural. Tested against facial recognition systems including Microsoft Azure's, the synthetic version of the same image and scene produces only about a 16% match to the original person, meaning it effectively cannot be re-identified, while the resulting data can still be used to train AI models for clients in automotive, data labeling, retail and smart cities.

Q&A highlights

Asked why Silicon Valley companies do not adopt privacy-by-design themselves, Strottner says major players are already engaging with privacy, citing one large Silicon Valley company opening a data-privacy office in Germany, and argues Europe's first-mover position with GDPR gives it a reputation and expertise advantage. Asked whether synthetic faces work for research beyond simple detection, such as emotion recognition, he says brighter AI's synthetic faces preserve the original's emotion, age and gender attributes without introducing artifacts, though use-case-specific testing is still needed. Asked about the biggest unresolved challenges ahead, he says the scale of data needed for self-driving and retail development makes privacy-compliant data a significant, ongoing problem that brighter AI is positioned to address.

Quotes

We see cameras as the richest sensor for innovative AI-based use cases.

Thomas Strottner

Today these systems can detect up to 20,000 key points in your face in order to track an individual.

Thomas Strottner

This is like often the perceived dilemma, and the question: which pill do you have to swallow?

Thomas Strottner

Questions answered in this talk

How does brighter AI's synthetic face technology protect privacy in camera data?

According to Thomas Strottner at the Rise of AI Summit 2020, brighter AI generates synthetic faces and license plates that look natural rather than pixelated, and in tests against Microsoft Azure's facial recognition the synthetic version of an image matched the original person only about 16% of the time, effectively preventing re-identification.

Why is pixelating faces and license plates a problem for training AI systems?

According to Thomas Strottner at the Rise of AI Summit 2020, pixelation introduces artifacts, such as a segmentation model seeing a phantom person or object, and can bias models, for example training a self-driving car system to expect pixelated license plates that do not exist in real-world driving.

How do privacy regulations like GDPR and California's Consumer Privacy Act affect AI companies using camera data?

According to Thomas Strottner at the Rise of AI Summit 2020, GDPR-style regulation already restricts how companies can collect public camera data in Europe, and California's new Consumer Privacy Act, modeled closely on GDPR, is creating a similar market in the US, which he sees as an opportunity for European privacy-centric solutions.

Mentioned

Facial recognition, GDPR, California Consumer Privacy Act, Federated learning, Synthetic data, Privacy by design, Data augmentation, Anonymization

Keep exploring

More from Rise of AI Summit 2020

More on Enterprise AI & Adoption

Transcript

Auto-generated transcript, not edited. · 3,497 words

Show the full transcript

we will have a talk about a very important topic something which is probably harder than many of us think um it is combining privacy and ai if it would be easy probably more people would be doing it um our speaker is Thomas Strottner he is business development manager at brighter AI which we just recently won the berlin deep tech award and he is filling in for maria anglais at the last minute so thank you very much uh thank you for filling in we're glad to have you welcome yeah thank you very much for the introduction um thank you very much um and as you mentioned um privacy indeed is is of course a very very important topic nowadays and in the previous talk that i loved it was a great talk um we also heard a lot about the dark about the darkness so now this is not only the talk of brighter AI which brings hopefully some brightness um but i will also tackle one of the biggest aspects or issues that we have

today which is data privacy and also how we can approach it um with an ai based approach to um yeah to solve this issue and to bring more uh brightness also to this world um so this is something i think we've all heard the quote we've all seen this kind of image what is very important for us to consider also as ai researchers is that for developing these kinds of systems you need data quantity a lot of data like these companies have but you also need good data quality and this is something that we will also get to a conclusion because privacy might seem to sometimes stand in the way here and so what we at brightly i focus on a lot and which is also very important for ai based innovation is camera data so we see cameras as the richest sensor for innovative ai based use cases so great examples for that is for example the amazon go shop shops where you have like automated checkouts all based on cameras um there are

a lot of use cases in smart cities also that start from people counting to a lot of other use cases and then of course they're self-driving cars where cameras also as tesla frequently is emphasizing um the cameras are actually the most important sensor here and that a lot of data needs to be collected needs to be processed in order to develop smart vehicles um then when we talk about the amazing achievements of ai um i think facial recognition can be one of them and of course most of us we really appreciate innovation that it brings like for example opening your iphone um just with your face and that's really amazing and we can see that in the last years with um yeah more like stronger modern neural networks um today these these systems can detect up to 20 000 key points in your face in order to track an individual and this is something that can be extremely useful and and um we should not focus only on the dark side of these things but

of course also on the positive however um if we look at recent news there's also of course a big risk of abuse and um i mean we are also an ai company and i'm also um yeah someone who who believes in the power of ai um but i think the the danger of having widespread surveillance and being monitored constantly as an individual their facial recognition is of course a big threat and we've seen this in the past so one of the big news was recently with a black lives matter protest that in fact facial recognition was already being used to track individuals we see that from also authoritarian regimes um where facial recognition is is used to to do things like social scoring um and there there's already a movement also by individuals that try to protect their privacy more strongly for example on social media as i'm showing here and there are more and more people yeah that feel that their privacy is invaded um and this is not only by individuals and by

society but we can also see that a lot of countries are actually acting on that um so um this map is showing um how data protection regulations are already in place and how strong they are and if we look at countries like the us and canada and european countries with the gdpr um then then data privacy regulation is already quite heavy and even in countries where sometimes you wouldn't expect it like for example here you see china also being marked as heavy um this is also an area where for example if a company an automotive wants to collect camera data in public um there are a lot of restrictions actually and it's not easy to um to collect and to process data that contains personal information um it's also quite interesting when looking at the u.s not many people have realized that in california this year there has been the introduction of the california consumer privacy act which is heavily close to gdpr and so a lot of people complain about our regulation but we

actually think this is also a chance for us to export privacy-centric solutions because california means it's not only the fifth largest economy if it would be a country um but it's also the home to um the the tech giants so this is quite an interesting movement also in san francisco for example facial recognition facial recognition technology has been banned so this is quite quite interesting to look also um to what's going on outside of europe now what we recently did was a study on the use of video data in particular and different countries we did this study together with the german ai association the kai bundesbaban and it was also emphasizing that there's more and more regulation in different countries but you know you see these large international companies like automotive companies um that operate very internationally but then you have regulation that can still be different in different countries so it's important to like pay close attention to all these laws and also to prepare for that and now what often happens is um

that there is a perceived ai privacy dilemma which we show here with like two different pills so sometimes you have on the on the left side here the powerful accurate ai models um but where there is a potential of breaking privacy so because you want to have very good structured data or you want to have very clean image and video data and this might lead to accurate ai models on the other hand if you need to anonymize data if you would for example blur faces and image data then you have worse quality and this is also something that we don't want um so this is like often the the perceived dilemma and the question which pill do you have to swallow um but there are actually lots of um yeah already potential solutions when looking at what can you do to emphasize privacy with different concepts so there's for example a different processing approach i will i will jump into this in a second and so there's processing that can be centralized but with more

privacy it can also be decentralized even for ai systems then when we look at data protection concepts there's for example privacy by default and there's privacy by design and then there's a last one that i would like to talk about which is um there's the option to alter the data itself and this can be through data augmentation or using synthetic data and as brighter AI we are somewhere here in between um so talking about synthetic data i think this was also in the talk before they are very interesting new technologies like generative adversarial networks um and then there are also new concepts like privacy by design and this is exactly the area where we are as a company so jumping first into the decentralized approach for more privacy this might be something that some of you have heard already in the past and there's a concept called federated learning i think this is also emphasized a lot by companies like apple um which mainly means that the algorithm is trained over multiple distributed devices or

servers that contain local data samples but not do not exchange them so this means for example in the in the example of apple that user data could stay on the end device and doesn't need to leave the end device but then a local ai model can also be trained directly on that device directly on the edge um what happens then is that these models would be encrypted and merged and this has also advantages when it comes to the use of resources and the resilience of the systems um however this is still in contrast to the more centralized approaches that are that are mainly used so most companies i was giving the example of automotive earlier they want to collect data and then work with this data in a central place so usually is the research center of the company then when we look at privacy by default this means that you would emphasize data protection with friendly default settings so you would ensure that personal data is processed with the highest privacy protection so that

by default personal data is made accessible to an indefinite number of entities um so this is like one of the things that also gdpr in europe is emphasizing and it means that by default you would have the most strict settings and in the real life this is something that we see a lot with cookie notices for example so by default these things should say you're not collecting cookies and then as soon as you accept something and you can also jump into a less privacy friendly setting then what you can also see is that you can use technology to create something that is called privacy by design so data protection can be most easily complied with if it's already fully technically integrated so that no subsequent measures are necessary um one of the ways that we know a lot is encryption using encrypted data in the first place and this is also something that's very important also these different measures are not either or but they should can also be combined and then something that we

can do is pseudonymize um attributes so this would mean for example that you would always use the same avatar for a person if you have image data of that person or you would always use a different name for a person in data set if it's medical data for example and then there's also anonymization where you would completely remove the data from the actually identifiable person so that it's no longer identifiable and as mentioned um like with adding privacy by design layers to your technologies you can also um yeah better fulfill requirements like gdpr um and it's also possible to combine these um and then something that that's very interesting um also if we look at the area of self-driving cars um is synthetic data um so here a fully synthetic image would be generated from scratch um and this can solve some of the problems but this can certainly not solve all problems i think um when you look at corner or edge cases for autonomous vehicles um this is inherent that there is a

problem because um our own creativity would probably never be enough to put all those scenarios in a synthetic data-like scenario but what you can do is you can also do like digital twins or create a virtual reality synthetic data where you don't have also privacy issues because you wouldn't have actual people you wouldn't have the actual car and the license plates but you have a synthetic world basically that you can use to train algorithms and then something that can also be used is data augmentation so with data augmentation you can use an existing data set and because at the beginning i said more data quantity but also quality is relevant so you can take data that you already have and with that data like for example here a clean image you could add certain artifacts to it so here for example you could add fog to the image and then you can create some of those corner cases by augmenting the data so you can have maybe a subset of privacy compliant data and then

augmented to actually increase the quantity of that data and also to have more more corner cases of course here again a simulation can never fully reflect the reality so um there can be a bias in that data um especially if you have the same scenes with different uh different augmentations like augmentation levels um so the the privacy dilemma um that is often perceived for ai when it comes to image and video data is something um that we know mainly because of this kind of pixelation so um if you look at a lot of studies or at the data that automotive companies are collecting for self-driving you wouldn't find a lot of pixelated data as a research scientist or as a data engineer you don't want to work with this kind of pixelated images and because what it does is it creates artifacts for your ai systems here on the right side this is a segmentation map by google d mines which segments all the objects and what you see here is that on the left

side the phantom person appears and also here in the head of that person there's a bird apparently and this is artifacts that are solely introduced to this image because of the pixelation um and at the same time you also don't want to train image images or ai systems based on this kind of pixelated data because it might increase the bias think about self-driving cars if you would pixelate all the license plates of the cars then maybe subconsciously you would train a neural network to actually recognize the pixelated license plate and if you would then use this kind of algorithm in real life it might happen that it's not detecting a car because in real life cars don't have pixelated license plates um so what we at brighter AI developed is we want to enable companies to collect the actual real data um but with privacy compliance so what we do is we generate synthetic objects for faces and for license plates so here on the left side you see a synthetic face here and this

is the actual real face and thereby this is a fully natural picture that you can use for your ai models and what is very important about that i was earlier showing how facial recognition models are really the biggest threat when it comes to identifying individuals in large data sets and what this can really exclude is the chance of re-identifying an individual and large data set so here we we test this with different facial recognition algorithms here for example from from microsoft azure and what this shows is that there's only a 16 match which effectively means this face belongs to a different person and this is even considering this that this image is the exact same scenery with the exact clothing and everything um so this is something um that is that is very fundamental where we as brighter AI work with a lot of companies uh in in automotive in data labeling in retail and smart cities um to take the pain away from data privacy that sometimes ai researchers feel at the same time

to have more data privacy um to remove this threat of re-identification and create high-quality data that can be used for ai systems so there was an introduction of of our company as as we had in the introduction this talk was supposed to be given by marian our ceo co-founder who couldn't be here on short notice um my email address is thomas at brighter AI my name is Thomas Strottner and thank you very much for listening to this talk yes thank you very much for the talk uh thanks again for filling in i have a couple of questions for you i encourage our audience to also ask further questions uh the first one uh would be you were talking about the california data privacy law which is is very similar and orients i think itself at uh at the gdpr and you said this might be very good for for you and maybe other european companies i was thinking um if this law exists in california why don't companies in the silicon valley do it themselves

or maybe the big players have already done it because of course google and apple they have to be active in companies where they're in countries where they already have to be gdpr compliant um so yes we definitely see it that these companies also engage in privacy topics so for example one of the uh big silicon valley companies open an office in germany um that is specialized in data privacy and i think this move also emphasizes one of the reasons why we are in a good position in europe to develop these kinds of solutions not only because we have a reputation but also because we have a lot of expertise in these areas um so we find this as a company that there's also um a lot of trust because as uh like with the introduction of gdpr europe the european union was the first one with very strict uh privacy laws um and and i think this first mover advantage we also have to take uh um for new solutions so i think we are

in an extremely good position um to be known for data privacy and to have a good reputation on that thank you and uh there's one more questions regarding the synthetic faces are these um are these pictures that you generate or that you modify good for all kinds of research and ai training or are they maybe incompatible when you try to train for emotions because obviously it's not the same face anymore yeah excellent question so this is this is in fact a different from use case to use case um so there might be there might be differences if you just simply want to do an instant segmentation you want to differentiate between objects like cars and people in general or you want to detect the emotions however in general we don't introduce any artifacts so the artificial faces that we generate have the same kinds of emotions the same attributes like age and gender that the original image has so um in general this can not only be used to augment your data sets but also

the attributes should stay the same um however if this should be tested so if there's anyone who's very interested and this needs to be tested on the individual data set and then we can also make comparison from how the system trained on the original data behaves and the system based on anonymized faces right so you're obviously developing edge technology i'm interested in what are the next big steps for you what are maybe big problems that you know you will have to tackle but are not so sure how to do it at the moment like where is the attitude that you perceive um so um i'm not sure if i understand your question correctly but um for us so the the problem of privacy and having good data for ai is actually not a small problem so i was mentioning some of the use cases and if you look at for example the amount of data that is being collected already for self-driving car development in the retail space this is quite a significant quite a

significant problem and so this is really something where we see an excellent fit and not only a big market but also something where we can really support companies in the future right thank you very much again thomas we don't have any more questions for you so uh again thank you uh we wish you a lot of success in your endeavors and we're very curious what brighter AI will do in the future thank you

Rise of AI 2027 · Berlin · May 2027

Be in the room for the next one.

Ten years of talks started as conversations in one curated room in Berlin. The next edition brings 300 selected leaders together in May 2027.