Video Transcript

Hello again, and welcome back to this second series, second episode in our series on privacy enhancing technologies. We kicked off the series with a look at Why Pets? Why Now? With Christy Chon Flynn and John Elliott. And in today's episode, we'll explore the adoption gap, why privacy enhancing technologies stall in practice with RSAC's Laura Ketzel and Natasha Gerla. Laura and Natasha, over to you.

Thanks very much. And Natasha, thanks so much for joining us today.

Yeah, thank you for having me to discuss one of my very favorite topics. Look forward to it. Excellent.

Just so people know who we are, I'm Laura and I'm the Head of Community Research at RSAC. And Natasha is the Director of Privacy and Data Policy at the Center for Information Policy Leadership in Brussels.

She does a lot of work on privacy and data related topics like cross border data flows, AI, data ethics, and data governance, and she's an experienced data privacy attorney. So I suspect that's why this is one of Natasha's favorite topics among other things. But before we start today, Casey mentioned that this is the second, episode in this series on privacy enhancing technologies. So if you were expecting pets as in cat videos, I am afraid you were in the wrong place.

Hopefully, though, you'll carry on because, you know, this is more interesting than cat videos. At least, that's awesome I think so. So I will briefly define pets for anyone new to the topic. And if you're interested in learning more about specific types of privacy enhancing technologies, go and listen to the first episode and check out the links to the related content that you'll see on this page.

Because today, what we're really gonna zero in on, as Casey said, is adoption and use cases and barriers to adoption for privacy enhancing technologies.

So here's that definition, and this is one that's used pretty widely, including by the European Union, which is a system of ICT measures protecting informational privacy by eliminating or minimizing personal data, thus preventing unnecessary or unwanted processing of personal data without losing the functionality of the system. So examples of pets, if you haven't listened to the first episode yet, are things like homomorphic encryption, differential privacy, and synthetic data. Hopefully, that's enough of a whirlwind foundation to be going on with so that we can start asking Natasha to talk us through our real topic here. Let's start out with some practical examples of privacy enhancing technology implementations that are really working and that are delivering real benefits? Because I think lots of people haven't necessarily encountered those too much in the wild. And Natasha, I know that you have. So talk to us a little bit about some of the best practical examples that you've seen.

Yeah, thank you very much. And I think that's really the right place to start the discussion because pets can sometimes sound a bit experimental or futuristic, or like you say, people think it's cats and dogs, when there are actually some very tangible use cases where pets are solving real problems today already. And one area I find particularly compelling is healthcare and collaborative research.

So if you imagine several hospitals that want to develop a better cancer detection model, for example, the obvious answer is to try and pool their pathology data, but that data is obviously highly sensitive and there are many good reasonslegal, privacy, security, or even commercial reasonswhy the hospitals cannot simply hand their underlying data sets to one another.

CIPL looked at such an example for one of our PETs papers. We wrote a couple of them, and I think, you know, they are freely available, and I may also be in the link later. And we looked at an example where federated learning was combined with a trusted execution environment, And instead of moving all the patient data into one central database, the model then would effectively go to the data. So each institution would train locally and the model updates could then be combined securely. The organizations, the hospitals, the research facilities gained the benefit of a much larger data set and a much more diverse data set, but they didn't have to create the risk of pooling all their data and pooling sensitive data with many restrictions on it together.

And I think that also illustrates something else that I find really important about pets. The long definition is privacy enhancing technologies, but I think that has become a little bit of a misnomer because they are, of course, privacy enhancing technology, but that's really only half the story because in my example, as you see, we also look at them as data enabling technologies and trust enabling technologies. They can enable the use of data that's beneficial that would otherwise perhaps not possibly happen at all.

And a second example I would give you is synthetic data.

We, for instance, have seen synthetic data used to train AI systems where obtaining sufficient real world data would be difficult or expensive or privacy invasive. One example, one case study we have in our papers involved generating millions of synthetic images of palms under different conditions to involve accuracy of a biometric system, for example.

And that allowed a model to encounter many more edge cases, training on the questions is always beneficial when you train an AI model.

It reduced the dependence on having to use real persons or real people's biometric data.

And there are many more everyday examples. You can think of differential privacy being used to learn population level patterns without learning things about you and me, the individuals involved in the population pattern, secure computation allowing, for example, two companies to discover useful information from their combined data sets without either one of them revealing, for instance, commercially sensitive information to one another. And a really prominent example is zero knowledge proofs in the context of age assurance.

So pets aren't really a single technology and they are also not necessarily something where you use one pet at a time. You may be able to combine, I call this a daisy chain of pets, you may be using them sequentially, for example, and they're also not a magic wand, they don't solve everything, but we already know that they can work.

And I think to the title of today's episode, what's interesting is a little bit the gap between that potential that we see and the actual deployment.

In a recent survey of organizations in European data driven marketing ecosystems, for example, three quarters of the respondents who use PETS were using just very familiar technologies such as anonymization or pseudonymization, which are, they do fall under this idea of pets, but then the use rate dropped to 16% only use differential privacy or 12% for trusted execution environments, and then the more complicated ones didn't have any replies, and the majority of respondents used them for something like GDPR compliance.

That's a very high adoption, but not unlocking data potential, for instance, that was less represented in the data set. So there's still a very real data between what pets can do and where organizations are usually using them today.

Got it. So that indicates that there are some stumbling blocks that those people are running into, or they just don't know what they don't know. So can you talk a little bit about what the barriers are to adopting a lot of these technologies that really might help people solve a lot of their common data sharing and data minimization problems and so on and so forth?

What do you think people typically run into and what recommendations do you have for getting past those barriers?

Yeah, thanks for that question. So we did look in our research, what drives the adoption of pets, who owns this, and what are the factors that go into this. One thing that came through quite clearly is that the successful pet deployment tends to start with a problem and not so much the technology. The question isn't usually, okay, we now bought homomorphic encryption. What can we possibly do with it?

You know, where do we use it?

But if we have two organizations that need to collaborate but neither can expose their data, then the question is, so how do we solve that? And is there something out there that could help us solve this? And once the business problem is clear and the business driver is clear, you can determine the appropriate technology. Quite often the answer then is, as I said, also a combination of technologies.

The second condition is cross functional development. That's a really important factor. So pets sit at the intersection of, you can think about this, like who drives this, who buys this, who has the budget for this. So the intersection is usually engineeringIT, privacy, security, product development, data governance, and business somewhere.

And so one of the barriers we found is that those groups can literally be speaking very different languages, also in how they define problems or what solutions they have. The lawyers may say, Okay, I need this data anonymized. The engineers need to know anonymized against whom or to what risk threshold for what use and with what utility remaining, because that's another question. If you anonymize fully, maybe the data isn't sufficiently usable anymore.

So there are trade offs with that. So successful implementation does require those teams to translate the legal or policy objectives or the business objectives into those technical objectives. And, research recently also identified that siloed responsibilities between legal, technical and business function, that's an important reason why pets that are available don't necessarily become part of an organization's broader privacy or data use strategy. And then there's an important point here on how companies approach pets.

They shouldn't be treated as a sort of specialist technology that gets bolted onto a system at the end. They work best when they are embedded into an organization's existing privacy by design or data governance processes, for instance, including for procurement or product development. That the question becomes at what point, at the point where we are designing this data use could a PET enable us to achieve the objectives with less exposure of personal data, for example. And then thirdly, there has to be a meaningful incentive for this to happen.

Sometimes that's access to data that otherwise isn't available, as I said, sometimes that is enhanced data security or data safety, and sometimes it's customer trust or it could be regulatory compliance. That last point matters because regulators have enormous powers to change the adoption equation here, right?

If an organization is considering investing money and engineering resources in pet technology, but has no idea whether the regulator will recognize that investment as meaningfully reducing its own risks or satisfies regulatory expectations, then the business case becomes a lot harder to make. So we have at CIPL, we have consistently said that pets should be really recognized by regulators as a demonstrable element of organisational accountability. If organisations go further than what the baseline requirements are and they invest in this kind of technology that demonstrably can reduce privacy or other risks, then the regulatory system should recognize that.

Got it. And I to ask a sort of more broader policy question here. So particularly here in Europe where both you and I are based, pets are sometimes lauded as a kind of possible silver bullet answer for both data sovereignty and digital sovereignty. So I'll define both of those just so that we we all know what we're talking about here. So data sovereignty is generally understood to mean that data on people and entities stays subject to the laws of the applicable jurisdiction, regardless of where that data goes. Lots of you were familiar with this in the context of regulations like the GDPR.

Digital sovereignty's definition seems to be a little more contested. So for the purposes of the discussion Natasha and I are about to have, let's use the European Commission's one, which is digital infrastructures, products, services that, in Europe's case, safeguard European security, strategic assets, and interests while respecting the body's values and requirements. So my question to you is, can pets really do all that? And if so, how?

Yeah, thank you. So, you know, I think maybe we broaden this a little bit in terms of what do, you know, as you said, what do we mean by sovereignty, particularly for The US audience?

Because this isn't just a question about the European GDPR again.

The sovereignty discussion is really increasingly about who can access data, who controls the infrastructure, where data and computation occur, which governments may potentially assert jurisdiction over that data, and how resilient is that infrastructure and how dependent is the organization or the governments on the particular technology provider, right? There are so many factors that flow into this. And with AI, that becomes even broader because we are increasingly talking not just about data sovereignty, but also control over AI infrastructure, for instance, and models and computing capacity and all that.

And so that's really part of what people mean when they talk about sovereign AI, the control not just over the data, but over the broader technology ecosystem in which the AI is developed and deployed. But I think we have sometimes collapsed all of these different dimensions into one very simple proxy. So where's the provider headquartered? Where's the server located?

And what does this mean for the jurisdiction? And that can lead very quickly to the conclusion that sovereignty becomes about localizing everything, the data, the infrastructure, the models, etcetera. The problem with that is, there are trade offs that come with this idea. Building and operating entirely local infrastructure can be very expensive, can be technically complex and skills intensive and can also mean you forgo some of the advantages of scale about security capabilities, computing resources, innovation, etcetera.

So this is where I think PETS can enter the scene and the discussion and can make the sovereignty discussion more interesting. They allow us to ask a more sophisticated question than simply, where's my data? But we can ask who actually has access to it and who controls that access?

Who can compute on this data? And that, you know, let's take confidential computing, you can design a cloud environment, for instance, in which data remains protected while it's being processed. And there are infrastructure operator, their ability to access the underlying data can be very significantly constrained in these contexts. With technology like secure multi party computation or homomorphic encryption, you can go further and where multiple parties can potentially collaborate or compute on very protected data without exposing it to one another.

So PETS potentially allow us to decouple collaboration from access, which is key in this sovereignty discussion. And it also means that maybe the choice isn't necessarily binary between keeping everything local and giving up some of the benefits of scale on the one hand or using a global infrastructure and giving up control on the other hand. And there's potentially much more interesting middle ground like hybrid architectures that combine something local or otherwise controlled infrastructure with privacy preserving computation.

And I think that's potentially a much more useful way to think about sovereignty, not as a technological isolation, but as the ability to participate with control in the ecosystem. Now, we have to be careful here because there isn't a magic wand, it's not, you set out pets and then you make the sovereignty concerns just vanish overnight.

It does mean that location or infrastructure location suddenly becomes fully irrelevant for all levels of data and pets will not substitute for law or strategy questions, but they can change the risk equation quite significantly, I think. And instead of treating sovereignty as this binary question, as I said, perhaps give us another layer of controls that can determine what is technically possible for another actor to do with the data we are trying to protect. That is where I think the delta between today's relatively limited deployment of pets and where they could be used may start to shrink a little bit if we think about it in this context.

Well, thank you so much, Natasha, for sharing your expertise with us today.

We really appreciate it.

Thank you for having me. There are certainly a lot more topics we can be discussing, and I have a lot more notes. Maybe we need to do Part B. But for today, thank you very much for having me.

We'll look forward to part b then. And everyone, please check back next month for the next installment in our privacy enhancing technology series.