#124 Sina Kian: Reshaping Privacy in the AI & Machine Learning Revolution

0:00When you think about machine learning or AI using databases, one of the most obvious questions is what's in those databases? And do we want to expose that to, in the context where I was talking about unaffiliated nodes, in this context to machine learning? You can use encryption to make data unavailable to the general public that may try to look at that data. But at the same time, if you put it on something like Elio, you can prove things about that data that are true or false or that fall within certain parameters and leverage that data without actually giving away identifying information. This week, I talk with Sina Kian, Chief Operating Officer of Alio, that's A-L-E-O, a platform for building private blockchain-based applications.

0:54Alio supports zero-knowledge proofs, a powerful tool to verify the correctness of a computation without revealing the input data or the computation itself. This makes Alio particularly useful in machine learning, where sensitive data such as medical records or financial information needs to be kept private, but has tremendous potential for training AI systems. Before we begin, I want to mention our sponsor NetSuite, Oracle's cloud-based enterprise resource planning software to help any business manage their financials, operations, and customer relationships in a single platform. For the first time in NetSuite's 22 years as the number one cloud financial system, you can defer payment on a full implementation for six months.

1:58No payment, no interest for six months for a full implementation of Oracle's NetSuite. To take advantage of this unprecedented financing offer, go to www.netsuite.com slash I on A I, that's E-Y-E-O-N-A-I, all run together. Now here's Sina. I've had a little bit of a look, but why don't we start by having you introduce yourself and then introduce Alio, and then I'll get into some questions. Okay, great. So my name is Sina Keehan. I'm general counsel and COO here at Alio. I graduated from law school. After law school, I clerked on the DC circuit for two years and then the Supreme Court for Chief Justice Roberts.

2:53I practiced law at a firm called WilmerHale here in DC for a couple of years. Then I changed gears, actually went into the finance world, worked at Blackstone for about four years. And during that time in an investment capacity, I got very interested in the technology behind blockchains generally. And I met the folks over here at Alio, got very interested in what they were doing. I joined initially as VP of strategy and have since been promoted to COO and general counsel. So happy to talk about Alio itself. As background, I was also an advisor to the Privacy and Civil Liberties Oversight Board.

3:32So long been very interested in issues around privacy and how they work, consistent with sort of national security and other public policy goals that we have. So Alio, if you're going to step back if you think about blockchains generally. If the internet allowed unaffiliated nodes, which is to say computers, phones, other nodes like that, to communicate with each other, even though they're unaffiliated and have no reason to trust each other, blockchain sort of took that a step further and allowed those unaffiliated nodes to cooperate on a task and specifically on updating a ledger. Now, the mechanism by which they did that is basically yelling the answer out loud, which is one node will say, hey, Craig sent Sina one Bitcoin or one ETH or one whatever, and all the other nodes write it down.

4:30And so that's how they update the ledger in terms of like data. Now, if you and I, Craig, started talking about what are use cases for ledgers, we'd start talking about obviously the financial use cases, bank accounts, that sort of thing, maybe real estate titles. We start talking about healthcare records, vaccine records, maybe voting databases. And for almost all of those use cases, if not all of them, you want some level of data confidentiality and privacy, which is to say you don't want to yell the answer to the entire world of unaffiliated nodes that you don't know and maybe don't have great reason to trust.

5:11And so what we're doing at Alio is we're allowing the same thing to happen, but for the answers to be encrypted, which is to say they yell the encrypted answers out loud. And you can still verify that everything being said is true. You just don't know what the actual content is as a completely unaffiliated node. As a user, as a program, you may, but in terms of what you broadcast to the entire world, that's encrypted. And that unlocks a lot more use cases. So we're very excited about that. So that's a very long answer to help explain what we're doing at Alio. Yeah, well, that's interesting. And you're specifically targeting machine learning solutions.

5:52Is that right? The overlap between what we're doing in machine learning and AI is really, really interesting because there are a couple of interfaces here. The first is training data. When you think about machine learning or AI using databases, one of the most obvious questions is what's in those databases and do we want to expose that to, in the context where I was talking about, unaffiliated nodes in this context to machine learning? And so one technique that you could use is something called homomorphic encryption. Another one is federated learning. Another one is actually exactly what we're doing, which is using a kind of cryptography called zero knowledge proofs to prove information from data without actually revealing the underlying data, which would allow machine learning to train without necessarily having exposure to the kinds of data we don't want, machines we don't quite understand yet to have access to.

6:54Yeah. Go ahead. Just to slow down a little bit, so I'm familiar with homomorphic encryption, and I'm familiar with federated learning. How is this different from either of those? So a zero-knowledge proof is basically an application of cryptography that allows a user to prove something about a set of facts without actually sharing that underlying set of facts. So it's a particular, and as a statement has three features. First, the proof has to be complete, meaning if someone provides the proof, we know with certainty that the underlying statement being proven is true. The second is it must be impossible for someone to provide a zero knowledge proof of a particular statement if that statement is false.

7:51And then finally, and this is the key one that distinguishes it from some of the other techniques we're talking about, is the proof must be zero knowledge, meaning the proof must not reveal anything about the statement other than the fact that the statement is true. And so homomorphic encryption overlaps with that very significantly, but it's its own distinct field. And in a machine learning training scenario, how does that work? I mean, federated learning, you're sending the model out to, different data pools that various owners have and you're training and then coming back and updating the model so that that data never has to leave the possession of the owner And the only thing that goes back are the weights from the training exercise.

9:03In this case, where is the model if you're training a model? And where is the data? And yeah, how does it work? Yeah. So, okay. Take a very simple example. example. Imagine that we are looking at certain kinds of health data and we want to know if the health data shows correlate or is there something we can learn about it based on age, which is people under 21, they have certain traits. Over 21, they have certain traits. Over 65, do they have certain traits?

9:39And in that context, you could actually look at, sorry, Can you hear baby in the background? Yeah, it's okay. It's okay. It's fine. So in that context, you can actually look at the health traits you care about and get proofs about the age of the person without actually knowing anything else about the person. So if you imagine a human doing it, they may pull out files. They may see, for example, things like the race, the name, the ethnicity, the gender, and they may allow that to affect their thinking. But if you're having like a machine learning context where the only thing you want to do is isolate sort of a scan result and the date of birth, then you can actually have it only prove that, okay, here are the people who are over 21 and here's what's happening.

10:30And is there a correlation or is there not without actually ever having to look at anything else? And so in a way, you can sort of think of it as helping to blind the learning so that it doesn't consider factors we don't want it to consider. Yeah. And the data in that case resides, say, in a hospital database? Well, so if you're talking about a blockchain, you can actually put it literally on a blockchain and then it can be leveraged by hospitals, insurance companies, anything, except they won't actually have access to the personal underlying information. So it's a matter of how you structure the data.

11:09You can have it traditionally sort of be on a database where it's like managed. In that case, you don't need a blockchain to accomplish anything I'm talking about at all. You can leverage something like$0.00 Crucible without a blockchain, or you can look at a world where you put it on a blockchain, it's completely encrypted, which means that no one can go to the blockchain and see your personal information or these health records, but you can, if you opt in, provide proofs of certain characteristics about it so that it can be studied. So it's a way of allowing your users to have a lot more sovereignty over their data.

11:44What can be exposed? How can it be exposed? And then you can combine that with machine learning to be more precise about what you're studying and what you're looking at when you're looking at data sets, as opposed to allowing machine learning to just go at it, come back with whatever it may come back with. Because what we know already is the data sets aren't totally neutral. They reflect human biases. And we may allow or permit machine learning algorithms to adopt those human biases without even understanding that we're doing that. So this allows us to adopt more safeguards against something like that.

12:20Think of it as using encryption to obscure what you're putting onto a database. The same way your passwords are obscured, same way all sorts of sensitive information is obscured, your credit card information, when you send it to the internet, you're not sending your literal credit card information. It hashes so that if your data packet is intercepted along the way, people don't actually have your credit card number. So similarly, you can use encryption to make data unavailable to the general public that may try to look at that data. But at the same time, if you put it on something like Elio, you can prove things about that data that are true or false or that fall within certain parameters and leverage that data without actually giving away identifying information.

13:11And you can tailor the parameters of that however you like. It's totally an open design space. And so this is very new. It's very, very, very early. and it's kind of ripe for creative thinking. Given that you have this tool that allows you to upload information, protect that information such that it's with encryption and yet be able to draw inferences from it and understand facts and correlations from it, how do you best leverage this? One of the interesting things, so there are two things going on here. There's the machine learning aspect and there's a blockchain aspect. What's interesting about the blockchain aspect is how it allows you to leverage something that doesn't have to live on one platform.

13:56So it's not just one hospital doing this. Every hospital can put their data together in something like this, and they can all benefit from all of the data that they're cumulatively providing without having to reveal anything about their patients to the world, without having to compromise their patients' privacy, and with getting their patients' consent of this. And they can allow their patients to retain ongoing sovereignty over this, which is to withdraw, do whatever they want, to own their data. And so thinking about how a system like that enables better forms of machine learning, I think is just a totally ripe design space, not only for products, but just philosophically.

14:40at the computer science level, at the level of theory, and at the level of actual practice. It's just totally new. And I think it's very, very exciting. And that's just on the training part. There's a lot more that helps with, and I sort of hinted at this, with actually like the output, how you can help control what factors are being looked at. And also, so take one of the products we're working on that we think is a good use case for our blockchain, which is digital identity. We allow, through this product, folks to be able to upload facts about their digital identity. It could be their passport.

15:24You can scan your passport onto a blockchain. And then you can leverage that passport going forward across platforms, completely agnostic to the platform. which is to say, right now, we all take for granted the status quo. If you go to Amazon, you give them your information. You go to Walmart, you give them your information. You go to some random website, you give them your information. But you could actually just have passport-grade digital identification that you leverage at each of these websites without actually having to reveal much more about yourself, depending on the website, depending on the nature of your relationship with them.

15:58Now, combine that with what's going on in something like machine learning, it doesn't have to just be your passport. It could be your vaccine records. It could be anything. And then you can allow machine learning to, you can allow a more competitive machine learning landscape, which is to say you can take that same data to multiple different algorithms and allow them all to train and try to come up with something from it, as opposed to having the data live on just Google or just Microsoft, where one company, We're sort of relying on one company to do its best with that. And it's a much more open source view of the world.

16:37It's a much more competitive view of the world. You know, one of the troubling things when you look at AI is how singular it seems to feel. You know, one or two or just a handful of companies potentially can own this space because of their access to data. But if you could open up that data in a way that is privacy preserving for the people whose data it is, then you can allow a lot more competition. And if you sort of believe classical theory, competition is going to result in better innovation, better products, and more responsiveness to the policy concerns that many have raised around AI. Are you familiar with Oasis Labs?

17:18I've heard of them, yeah. Yeah, because it sounds similar. They have a combination of secure hardware and cryptographic techniques to allow privacy preserving computation on the blockchain. But they have kind of a blend of strategies. strategies so from for alio's uh users uh uh all of this happens through a web interface or through an API and uh does uh yeah explain how it works from the from yeah what we're building is an open source protocol that we're hoping to launch later this year when this open source protocol launches, our audience for that is actually developers. It's people who build the applications we're talking about.

18:23We're not, for example, building a healthcare application at the moment, but I believe as people learn about what we're building, they'll start to see the potential for the things that we're talking about, which is, okay, you can leverage this open source blockchain and you can build an application on top of it that uses, for example, It funnels certain kinds of data, puts it on this blockchain in a highly secured way, and then allows it to be leveraged going forward. And one of the reasons I came up with healthcare is everyone kind of intuitively understands there's something wrong with the way healthcare records work.

18:59You have your primary care physician when you're younger. Maybe you move. Maybe they retire. It's very hard to get your records. A lot of people don't know if they were vaccinated, when they were vaccinated, what they were vaccinated against. in that environment to actually get together, okay, who is everyone who's had a screen for colon cancer? What are our accuracy rates? Who's doing it better? It's very hard because the data is so fragmented. And one of the reasons it's good for the data to be fragmented is because it's privacy preserving. If all the doctors just said, hey, everyone, here's the scan for my latest patient, then your information would be sort of revealed to the whole world.

19:39And that's not an attractive sort of way of creating machine learning, enabling data. But what I'm talking about, if there are developers out there listening to this and they're thinking, oh, that's my wheelhouse, they're our audience because they should be building, experimenting with how to do this. And when I say this, I mean, create an application that allows you to upload this specific kind of data or any kind of help data allows you to leverage it and gives you the benefit of access to the data without actually compromising any individual's personal information. And our laws aren't completely foreign to this.

20:23Our census law, for example, allows the Census Bureau to analyze data and to publish reports on that data so long as it's functionally anonymizing, which is to say the nature of the publication doesn't give away. It's not so specific that it would give away who it's talking about, right? So collecting data and preserving its privacy while benefiting from it is not foreign to us. This is just a totally new tool that I think is going to be really enhancing for the AI machine learning space. Yeah, and it's open source. So the business plan for Alio is to open source it, get it widely adopted, and then offer sort of expert services or consulting on specific applications.

21:18uh applications exactly right yeah yeah who's the competition i mean are there other open source projects like this i know this has been an active space i mentioned oasis labs uh i've had don song uh who's the the berkeley professor behind it on the podcast uh but i frankly haven't looked at any of this for a long time. Do you, is there a competition? First of all, I think I would do terrible injustice to all the really talented people out there working on this from one angle or another. So I would be reluctant to sort of have a list. Right now, I don't think of it as a competitive space. What do I mean by that?

22:02I sort of think of it as an innovative space where everyone's trying to build out what the technology can enable at an infrastructure level. and then what's enabled on top of that is really, really incredible. And then that's sort of when competition kicks in. So now when I see people working in this space, I sort of view them as really important in a sense partners because they're doing R &D, they're doing the marketing. And the reality is there are computer scientists, very smart computer scientists like Dan Bonet over at Stanford who know this inside and out. But the world of developers, they're still learning about this.

22:38And so anyone kind of introducing these tools to the world, explaining these are tools you can use, I view as sort of a partner. It's just too early to think of. It's sort of being like, who are the competitors, you know, around the Internet in the 1980s? Like you could conceptualize competitors, but it's just a little too early to think of the space that way.

23:01What is ZKML I see on your website? Yeah, so ZK in our space is a common way of saying zero knowledge. Okay. So it's zero knowledge machine learning, which is to say machine learning based on an algorithm that does not have access to the underlying data, can only draw inferences from it. Yeah. So where are you in the project right now? You say that it has not been launched or, yeah. Yeah, we're in the midst of our test at three. And so we just opened up the ability to deploy applications, and we've had many hundreds of applications deployed on our testnet. So far, it's working really well. The point of our testnet is to identify bugs, to invite an open source community to sort of begin the project of making this a global open source project, as opposed to something that we own or are particularly important for.

24:05because, as you said, we'd like to pivot to applications and enabling use cases on top of it. Yeah, and how large is your developer base right now, people that are participating? Yeah, there are lots of different ways to measure it. There's the programs that are being deployed. There's the nodes that are being run, many thousands. Right now, what I find heartening is that there's a very enthusiastic core group of folks thinking about zero knowledge on our testnet, on other projects. And that group seems to be getting larger and the contributors are increasing in what feels to be exponential with exponential numbers.

25:00So we're sort of very heartened by all the trends we're seeing there. And then the number of people who are not computer scientists who are familiar with zero knowledge proofs is increasing. The Treasury Department released a report referencing them. The Office of Science and Technology Policy at the White House is getting familiar with this technology. So the audience for this is growing really, really fast. And we're very excited about that. And I think it's time for the machine learning world to recognize the Venn diagram overlap here because it's very interesting. I mean, let me give you another way that this can be leveraged that I think intersects with machine learning and the user experience online.

25:46So right now, if you go to Amazon or Walmart from the United States, through your IP address, roughly knows not only you're in the United States, but like maybe what city you're in and maybe even more specific than that. And you don't think too much about it, but the products that it shows you are a function of that. So what's available to you given where you are. And if you went to France, you'd have a slightly different experience. And if you went to another country, maybe you'd have a very different experience or no experience at all. So that's to say that the internet currently, and by extension, any machine learning model on top of that, understands where you are and gives you an experience that's somewhat tailored to that.

26:26We could also build it so that, let's just take the digital identity concept. Amazon or Walmart or the next website would know that you're under 12 or under 18 or over 21. And you wouldn't think too much about it, but what's available changes. Is alcohol on the website? Is certain adult content on the website? Or dangerous things on the website? You know, things that 12-year-olds should be playing with on the website. And as a result, the machine learning that's built on top of that could also be leveraging that information, again, without knowing name, date of birth, anything specific to the individual, just that, hey, maybe certain things are age appropriate.

27:12Maybe we don't want machine learning for children of a certain age. Or maybe if we do want it, we want it tailored a certain way and we have certain public policy goals that matter. So having credentials like this that can be leveraged and that the entire internet would be responsive to has the trickle-down effect of any sort of what I'll broadly call machine learning on top of that, sort of watching the user experience, seeing how the user interacts would be sensitive to that particular criteria. And so you can imagine when you're thinking about TikTok or social media, that this could be very important, especially to parents.

27:54Yeah, it's, so that application relies on kind of the passport idea, the ID for an individual that then is accessed by whoever

28:15wants to access that individual? I mean, how would that work? Yeah, so think about how it currently works, which is you go to buy wine on a website, they ask you, are you 21? You click yes, that's the end of the questioning, which probably good enough security for wine. We probably don't have an epidemic of minors ordering wine on websites. But where you have greater concerns, maybe you want something more than just data that an individual can arbitrarily input. You know, if they know they just need to pick something that says they're over 18 or over 21, they'll just pick it. And so you can have a lot of fake identities.

28:50I mean, this applies in so many contexts, by the way, Craig. If you look over at, like, ticket sales, something like 40 % of ticket sales are sold to bots. So put aside a human lying about itself. It's just like something that's trained to lie to the actual merchant. You can replace that with something where you have, again, passport grade identification, which can't be faked, digitally signed by the U.S. government, can be authenticated against the U.S. government database. and that would be the identifying credential. And so the ability of a child to suggest they're over 21 or the ability of someone to suggest they're a US citizen when in fact they're a North Korean citizen, I think it limits that very significantly because you're talking about passport grade security where some of these websites just have clicked through, you're 21.

29:47and and to to develop that uh and that that uh there would be a key or something that yes someone would use there'd be a combination of a of a private key which is specifically the individual uh and live scan which is to say something to so if i had your private key uh you know the live scan would show actually it's not craig even though he has craig's private key uh and in the case of parents, you can imagine introducing multi-party computation or other things so that maybe there are even two or three keys are necessary for login. It's just a matter of how you, again, it's an open design space.