EP121 What Happens Here Stays Here: Confidential City (and Space)

Security.

Hi there. Welcome to the Cloud Security Podcast by Google. Thanks for joining us today. Your hosts here are myself, Timothy Peacock, the Senior Product Manager for Threat Detection here at Google Cloud, and Anton Chuvakin, a reformed analyst and senior staff in Google Cloud's Office of the CISO. You can find and subscribe to this podcast wherever you get your podcasts, as well as at our website, cloud.google.com/security. If you enjoy our content and want it delivered to you piping hot every Monday, please do hit that subscribe button. You can follow the show and argue with us and the rest of the Cloud Security Podcast listeners on LinkedIn as well. Anton, we have once again welcomed Nelly, and this time Renee, back to the podcast to talk about perennial favorite topic, confidential computing.

Yes, we did. And I think that perennial favorite is frankly pushing it. It's a topic we go and visit every year, like an old relative.

I think I think this was one of the more fun episodes to record because— so listeners, I was a debater in college and every weekend I would go somewhere and as a competitive sport, argue with people. And I've learned in life now, it's not nice. It means that I like to play rough. And on this episode, we actually had the opportunity to get into it a little bit, have a little bit of a tussle. And I think we all came out the other side still friends. What do you think, Anton? Was it acceptable?

I think so too.

My thoughts went into a different direction because I was like, when you said, and then you learned that it's not nice, I had the oh, shh moment of like—

Oh, was that me speaking today?

Yeah, yeah. It was almost like, oh no, I did miss this particular memo and I actually haven't learned it yet. So should I learn it? I started thinking about it and went on this whole soul-searching routine while you were finishing.

Well, listeners, I'll leave it to you all to imagine in your head how badly I want to say something about old dogs and new tricks right now. Oh wait, I did say it.

We'll figure out whether it's a worthy learning for me in the next 10, 20, 30, 40, 50 years.

But this was a really fun episode.

Yes, it was a very fun episode.

We talk about use cases, we talk about threat models, we get into a little bit of like really pushing on where does confidential come in, what are the alternatives, and why did we go down this direction? So listeners, if you're interested in how a really, really highly esteemed product leader at Google and security leader Nellie thought through these problems as she brought something to market, as well as Renee's contributions, I think this is going to be a really fun episode. So with that, let's welcome today's guest. I'm delighted, delighted to introduce today's guests. We're joined today by Renee Kolga, a PM for Confidential Compute here at Google Cloud, And returning podcast guest and legend, Nellie Porter, Group Product Manager for Encryption and Confidential. Welcome back to the show. We are delighted to be talking about confidential compute once again. Could you, one of you two, and this isn't for either of you, give us the 90-second overview of what confidential is and maybe contrast it with some related things like k-anonymity in, say, a data clean room? What does it do? What doesn't it do?

First of all, thank you so much for having us on your podcast.

Absolutely.

Both of your team and Anton, and we're absolutely delighted to be here and talk to you. Renaud, please start.

Yeah, super quick. Great to be here, folks. But confidential computing basically helps complete the end-to-end data protection story, right? The encryption of data at rest and in transit has been around and well-known. But what about the data in use? And prior to confidential computing, there wasn't really a way to protect data, those ML models while they're being used or processed. So basically confidential computing strengthens this data protection story through hardware-powered encryption of memory. How about that?

Rene, and again, we have those 2 additional questions coming from Tim. Sorry, Tim, interrupting.

Well, Nelly, why don't we make those your questions then? How do you contrast it with, uh, Yeah, with, say, like k-anonymity, because that also gives us protection about data in use, right?

Absolutely. So anonymity, or ability to anonymize data in some different way, shape, and form, one of technique to make sensitive data less sensitive. So it's very common mechanisms that a lot of technology using. The problem with this approach, again, the value of this approach is making your sensitive data less problematic. So you can share, you can do something about that. But problem is that a lot of, again, ML and AI and again, big bots in the world, as you can imagine, absolutely need very high fidelity of the data. And as more noise we introduce in this data, is less and less for us to be able to get high fidelity results. So I think all of those technologies have significant value to the market of data protection, but confidential computing, I believe, provides this opportunity to not introduce too much noise and keep data naked, completely clean, and being able to do a lot of work on this data in those strict environments so you would not worry about keeping data in full fidelity. And then the result, especially when we're talking about AI/ML, is going to be so much better, I can promise you.

Yes. So basically, to kind of summarize, confidential computing allows for higher precision data matching or higher quality ML training because you don't have to anonymize your sensitive data. At the same time, you can certainly combine these techniques together as well.

Oh, and one additional thing that I forgot to mention. Thank you so much, Renee, for summarizing. It also allows us to run general compute. Imagine, and what we've seen so far, there are so many different algorithms that need to be processed in this environment, and being able to not force any limitation and allow general compute to proceed in these cryptographically isolated spaces It's absolutely values that I think confidential computing can bring to the table. Back to Rene.

No, Nelly, it's our job to pass between guests. And I think it's actually on Anton to ask another question.

Yeah, I think that I don't wanna prod too hard, but you mentioned something.

No, prod, prod, prod, prod. Do it, do it, do it.

I'm gonna prod, but I won't prod too hard. That's what I said. I didn't say I'm not gonna prod.

I'm gonna prod, just not too hard.

Prod politely.

So, for example, ML training where we are talking about confidential versus, say, anonymized data, that to me makes sense. But ultimately, you can just train it on private data and kind of like take a risk, right? That's also an option for some people. So, what threats does this really stop? And I'm getting a minor déjà vu that during the previous 2 episodes, we've asked nearly the same question. Hey, confidential sounds amazing. What threats does it stop? And are these common at clients? What types of clients?

Yeah.

So give us another look for 2023. Threats, compliance mandates, what is the actual problem? What is the badness we are solving with this?

Or maybe to put it another way, if we use confidential, what gets better?

Yes, that's actually a really good question you learned from somebody on the show. Me.

I made up that question on the show.

Oh, right. You made it up.

Good.

Thank you so much, Tim. So let me start because, again, I was trying to answer the same question multiple times on the same show. So let's continue this trend. So you will compare and contrast, and again, embarrass me after the show, and I will let Renee to help me with that. But let's start with compliance. There is no, again, maybe specific regulations that would demand specific type of protection, and it's never been because all and every regulation, every compliance regime is actually trying to be a little more abstract on the intent and how to implement it is up to, again, somebody that have to meet those regulations. And what we've seen more and more as we move into 2023, customers see confidential computing as a helpful additional control to implement the protection. We're talking about HIPAA, we're talking about GDPR, we're talking about very many other regulations that come into play. One additional very interesting development that we also notice, a lot of focus on hardware-based security. Literally a week ago, NIST published a very interesting paper recommending higher level of protection, especially in those places where data is very important and data protection is very important. And they're talking about Confidential computing, HSMs, and all other mechanisms how hardware-based security can help us to raise the bar and get more protection to our customers. So, it's kind of the compliance story.

So, yeah, this time I kind of feel like I need to prod because you've said 2 things that in my experience tend to trigger CISOs. And I'm not a CISO, but in this case, I'm kind of triggered as well. When many CISOs hear helpful additional control, the first reaction is, I ain't got no money for it. What are you talking about? I'm barely surviving with these ransomware fights or whatever. So, helpful additional control to me is kind of famous last words before you hear no to your budget requests. So, I think that the making compliance faster, cheaper, less onerous, I'd buy.

Yeah.

But helpful additional control screams, don't buy me.

Screams P2 to me.

Screams P2. Thank you for PMing it up, team.

Yeah, yeah. And P2 is my way of saying, that's a nice idea. We'll do it next of never.

Good. Not what we heard from our customers.

That's good.

Because our customers trying to meet those requirements, again, in the full intent. For example, we have customers that are saying, we can bring the sensitive data It's not only, again, compliance in the normal sense, it's also requirements that are coming from sovereign space. We can't bring this data outside of our country or outside of our region only when it's end-to-end encrypted. When customers think end-to-end encrypted, for them, it's the way how to use encryption to limit access and again provide this proper audit recommendation to their, again, compliance officer to provide the whole proof that they're doing what needs to be done in this case. So I agree with you, it's not optional in very many customers' minds, but it's still, again, emerging technology. It's still future.

Mm-hmm.

And we still have really serious work to do to explain the value and again, answer Anton's question about what specific value it provides. And I think if I will summarize it, forget about it for compliance only for a moment, this technology together and combined with all other technologies that we have, that's why I own encryption and confidential computing, end-to-end helping us—

I thought it was because you liked math.

I love math, by the way. Love, love, love. But it's not only that. It's also because we have opportunity to provide this end-to-end protection for our customers and eliminate unauthorized access as a result of that. And it's very critical part of our stories that we continue to innovate how we will tell it, how we will help customers to understand. Rene, back to you.

Yeah, I mean, just to second and add to that, it's, you know, that meme of, what if I told you that cloud is just someone else's computer, right?

Oh, okay, okay, okay.

Oh no, no, no, you're gonna get banned from a podcast. On the landmine. Mid-episode.

You have stepped in it, my friend. So listen, listen, listeners, you might be, why is Tim having this reaction? And it's because the cloud is so much more than someone else's computer. Rene's perspective that he was about to share, he shared this earlier at a company all-hands, and it was all I could do to stay quiet on the all-hands, was he was going to say, what if I told you the cloud is just someone else's computer? And I'm here to tell you the cloud is so much more than someone else's computer. It's IAM systems, It's serverless systems, it's managed services, it's AI, it's Kubernetes. There are a gajillion things in the cloud that are not someone else's computer. So, Rene, me having thoroughly done what I just did, please, please go ahead.

Yes, yes. But what the customers want is to give them control back, right? How can the cloud providers give customers technical guarantees that Even though all of this goodness does not belong to a customer, but the customer controls what the cloud provider can do with their data. What can they access? What they cannot access. And that's what that ties with the regulation, especially for European customers that say, hey, give me technical guarantees that only I can access my own data. And that's what confidential computing helps you with. And then on top of that, On top of what Nelly mentioned, it's about some of the more advanced attacks, like what if there is a zero-day in the hypervisor and somehow one VM can peer into memory of another VM? And again, confidential computing, as it keeps that memory encrypted, even if there was such an attack, such a zero-day, all what they would get is a bunch of encrypted data.

And I think the same line, if you say it, Cloud involves computers owned by somebody else, just doesn't have the same striking power. And of course, if you say this, neither me nor Tim would get, you know, excited, anxious. But it's just the line just doesn't sound the same. Cloud involves computers owned by somebody else.

Now when you say it like that.

And now it's exactly the point is that the space that we operate security is also not simply, again, end-to-end protection. There are so many different tools and capabilities that have to come together as a set of controls, exactly as Renee said, so customers will have transparency and control on how their sensitive data processed.

I think what's really interesting in there is cloud is totally someone else's computer. That's undeniably true. But I think what's really interesting thinking about this is what confidential gives you is a case where the key doesn't belong to the same person as the computer. The key literally belongs to the silicon, and you can prove that through and through. And so I guess what's really cool is you're not creating a world where the cloud isn't someone else's computer. You're creating a world where the computer and the keys are not one-to-one. And that is actually a really fascinating thing. So I want to, I want to shift gears a little bit and talk about something that we can finally talk about externally that I think you teased on each of your previous appearances on the show, Confidential Space. Could you give our listeners a little bit of an overview of what this is, why it's so cool? Because, you know, I might be being hard today, but I really think this is cool.

Tim, please be hard on us. It's absolutely okay. It's allowed. Yeah.

So there's been lots of innovation, and I'm sure Nelly can add more on what has happened in our overall confidential computing portfolio. But yes, one of the most exciting things, probably because I own that product, is a product called Confidential Space. Its internal name, was known as Las Vegas, and I think it was very appropriate because as the saying goes, anything that happens in Vegas stays in Vegas. And that's exactly what Confidential Space allows our customers to do. It basically creates a secure enclave or a trusted execution environment or a TEE, and it unlocks some super cool use cases that were previously really impossible where It's all about secure multi-party collaboration where multiple teams within the same organization or multiple external organizations can perform computation on super sensitive data together without sharing that data with each other.

So that's really cool. Hang on, hang on. How can that be?

What do you mean, how can it be?

Because math, right?

How can you do that? How can you let people do compute on data without sharing the data? How does that work?

Exactly, right? It's like sharing data without sharing data. You share data with some code, with some workload that you have together maybe created or you have agreed upon somehow. And the data is only seen by the code, but not by the collaborators. So say, imagine 2 banks, they want to find some fraudulent activity, money laundering, whatever, right? They compete with each other. Data is super sensitive, PCI, PII, who knows what else, GDPR maybe. And what if they could agree on some code that would know that the data needs to come in in this format, and then we look for similar names who took out loans or transferred money more than X, whatever, and provide the output. So, um, as a result, they performed mutually beneficial calculation, whether it was a ML model or simple query, whatever. They benefited from the output, but they didn't lose control of any of that data. They still maintained full control over the data. They only shared it with the workload and they benefited from the calculation.

So do they need confidential computing for that? Or would like zero-knowledge proofs and just agreeing on auditing the software ahead of time be enough? Why is confidential necessary to make this happen?

Well, it's to get those guarantees that what you expect to be running in that environment is indeed what is running in there. So we, on top of regular vanilla confidential compute, so to speak, we add 2 additional things in confidential space.

Sprinkles in hot cocoa?

Almost, almost. Better. We add a hardened OS image that removes things like SSH and serial, et cetera. And that locks out the admin or the operator of the environment from being able to influence the workload, access the data, et cetera. And then we add an attestation, remote attestation service that is able to give strong cryptographic proof of what is running in that trusted execution environment. Is it running on confidential VM, keeping memory encrypted? Is it running on that hardened OS image? Okay. And what code is it running? Is it image digest or hash 256 equals ABC123?

So is this like binary authorization conceptually? It's giving us—

On steroids. On steroids.

Nice. Got it. Nelly, please.

I think if I will come back to your question, Tim, and it's again because I'm so much in love with math, zero-proof and differential privacy. There is multiple mechanisms that allow us to do this multi-party computation. And as mathematicians, and all of us love that, it's a very sophisticated mechanism to accomplish a very similar goal. The problem, as usual, sorry, starting with problem, it's not general compute.

Hmm.

And it's very tied to specific queries or specific implementation of your algorithm. What we found, and it's exactly what Renee is trying to explain, we have so many different variations of what people need to do together.

We need something more general.

Absolutely. And this, again, sometimes people asking, okay, how we come up to that? And the story goes that we so much were focusing on strictly, strongly, cryptographically isolate people from each other when they realize that it's fantastic. Now they sitting on those islands with tons of data, and now they need to work together. What the— how they would be able to do it? And that's where it's coming from, that now when we have this, again, enormous amount, the whole lakes of data, how we would be able to bring them? And that's why cloud is critical, why this general type of environments to use is critical, to give people ability to work together and collaborate securely without leaking the data and without exposing this data. Why confidentiality important? Because it's confidential data or confidential code or both. And it's very— we can't survive without this capability. Yes, Anton.

Yeah, so maybe you can give me a kind of a differential view, or a diff rather, not a differential view, between the confidential computing before and confidential space now. So, what is the most important part of the new different things? What is the— what makes it a space rather than just a confidential VM or container?

Let me try to put it together. So, I think for confidential computing in general, the capability we provide, we provide confidentiality and some type of integrity within your environment, and we're protecting our customer workloads and our customers' data against their, again, multi-tenant presence, isolating them between tenants and isolating them against cloud provider and our zero days. What confidential space brings to place is absolute interesting capability of protect against workload owner. You're paying for that. But you cannot influence what is happening.

A workload owner. Wait a second. Not the cloud provider, not the next tenant, but against the—

A workload owner.

In addition.

Why didn't you say it? Like, I've been trying to get to that and finally I hear it.

Sorry.

Okay.

Sorry, Anton. So, and it means, and why it's so critically important, because when we're talking about multi- party. None of parties actually trust each other. It means if somebody is nominated to run the whole pipeline and this workload, they cannot cheat. They cannot get access to the sensitive data of everybody else. So the one of the paramount, again, experiences and security we have to offer protection against workload owner themselves. And it's what difference between confidential space and how we accomplish it. We made this system immutable and we made sure that we will not start those boxes until your container or your application and attestation integrity assurances of this application and the whole entire system will be validated. And only that, we will start this confidential space for you. It means even you admin, you would be able, congratulations to you, to start this box and stop this box. Yep.

That's it.

And it's exactly what needs to be done for multi-party secure collaboration space. Yeah.

And that's what this hardened VM accomplishes, that locks out the operator or the admin of the environment. So indeed they can only start and stop the workload, but that's about it.

So give me a couple of use cases that this necessitates.

So I mentioned the fintech or financial services example, right?

This one's really interesting to me. I like this use case.

Yeah, yeah, yeah. We have one of the fintech startups, Monetago, actually using confidential computing, and they are offering an ability for banks to detect duplicate financing, which is an interesting type of fraud. I didn't know it existed, But it's where one organization goes to multiple banks and takes out a loan against the same collateral, like a future cash flow or some future deliverables. And obviously, banks don't really like that because if this organization goes bankrupt, they won't be able to collect on their loan. Insurance space has double-dipping fraud where, again, same accident is claimed across multiple insurance agencies. They don't like that either. But beyond that, we believe healthcare organizations, pharma, biotech, they have a lot of sensitive PHI data that they would love to analyze and pull together and train some amazing ML models. We have Web3 organizations and they want to improve the way they store secrets and how to better and faster and more securely sign those digital asset transactions. And ad tech, frankly, is probably one of the largest areas that needs TEEs, right? Because third-party cookies are going away. That's great for privacy, but how do we still allow this free and open web continue? And how can we provide advertising in a privacy-safe way? So there are a number of initiatives out there, and many of them require the use of TEs for like data clean room, secure data matching, etc. Nelly, anything else?

I think you covered most of the cases. I would say that we've seen again the usual and different sectors from usual, again, digital native customers to more financial customers that actually very still in this transition to cloud. This capability allows them to address some of the biggest issues when they need to collaborate effectively and securely across boundaries. And sometimes in organizations, they have these Chinese walls within the organization itself that cannot communicate and, again, collaborate effectively. And being able to have this safe space, secure space, where it can happened allow enormous amount of very interesting use cases. As it, again, Brianna mentioned a few of them.

Well, I hate to bring us to a close because this has been one of the most dynamic episodes we've recorded in a long time, but we all have real jobs to get back to. I wanna ask our traditional closing questions to both of you. One, do you have recommended reading? And 2, do you have a tip for improving the confidentiality of our listeners' data?

Okay. What's next and how to improve? Well, of course, not being super biased here, I hope our listeners give confidential computing a try because we really made it like super insanely easy to try. Like literally you click one button and your workload is running in confidential computing. It's that easy. So you click the button and off you go. So give it a try. reading. I'm in a very positive mood today and it's not confidential computing specific. So I'll recommend a book called This Is How They Tell Me the World Ends: The Cyber Weapons Arms Race. I don't know if you read that.

Ah, that's a great book.

By Nicole Perlroth.

Yeah, she's excellent.

Yes. Yeah. So I found it slightly repetitive, but I really love the kind of the whole zero-day marketplace and sort of Super interesting.

That's awesome. Nelly, same questions, please.

I would say think about hardware. We strongly believe that hardware is the way to go to provide high-level guarantees for data security. And when we're talking about that, my recommended reading— sorry, being incredibly boring in this particular respect— this NIST paper is actually very interesting reading. We can share it with you, but it's very interesting reading that coming from many, many parts of industry, and it's providing really interesting way for customers to assess their understanding of different types of technology and understanding how it's falling together. Because we have this kind of completely isolated walls that we usually trying to isolate even more. It's time for us to come to this safe place, confidential, I hope, where we can combine this knowledge and put it together. This NIST paper is also an attempt to connect the dots across so many interesting innovative technologies that will help us to bring data protection to the next level.

And given that it's a NIST paper, it's probably only maybe 400, 500 pages, so it's pretty short.

Well, that's bedtime reading It will be like—

Just 1 or 2 months of reading and you're done. It's very quick.

It will be like book.

Bedtime reading.

Nelly, Renee, thank you so much for joining us today. I hope you enjoyed this as much as I did.

Totally. Thanks, Tim. Thanks, Anton.

Absolutely. Thank you so much.

And now we are at time. Thank you very much for listening and of course for subscribing. You can find this podcast at Google Podcasts, Apple Podcasts, Spotify, or wherever else you get your podcasts. Also, you can find us at our website cloud.withgoogle.com/cloudsecurity/podcast. Please subscribe so that you don't miss episodes. You can follow us on Twitter, twitter.com/cloudsecpodcast. Your hosts are also on Twitter at Anton_Chuvaikin and Tim Pico. Tweet at us, email us, argue with us, and if you like or hate what we hear, we can invite you to the next episode. See you on the next Cloud Security Podcast episode.

Security.

Security.