What does a encrypted supercomputer look like? Yannik Schrade [Arcium] & Robert De Groot in dialogue

first of all yanic thank you for making it all the way out here thank you a lot um yeah very nice venue actually can be a bit searching where to go but eventually we'll end up all here so um thank you all for joining us today here I'm perhaps um I'd like to kind of like to ask you to give a short introduction about who is the man sitting here more importantly what have you been doing with arim um this is not a shilling panel this is just I'm interested in what you guys do with NPC and what it means for the industry and what are the trends as a whole with that with regards to the topic of today so yeah sure so um at arum we are building a decentralized confidential Computing network with secure multiparty computation so the way you can think of arum is not as a blockchain network but rather a dedicated purely computational Network that allows for confidential computations to be executed on existing blockchain networks um and um yeah the the way I'd like to think of it as really arum being it's dedicated encrypted supercomputer and then existing blockchain networks being IO devices to that to that computer and uh with this cognitive model it also makes sense to then have offchain sources run confidential computations without actually having to to care about blockchain applications as a whole and um one of the connections where blockchain really comes into play with with arum is um with a sensorship resistance mechanism that we've designed so um with arum it's possible to run any kind of confidential computation arbitrary confidential computations over encrypted data um without having to to expose that data or share that data um and yeah by combining blockchains with with MPC and um staking and slashing mechanisms it becomes possible to have um the lowest possible trust assumptions with high efficiency um and sensorship resistance could you give for the audience perhaps an insight into what kind of data sets would be interesting what kind of uh verticals would be touched by this uh how would this have an impact sure so um I think there's there's different kinds of of applications being built on top of arum um the most straightforward kinds of applications really are onchain applications so on an existing public Ledger someone wants to build a um confidential order book for example wants to build some confidentiality into their smart contracts um and with with our architecture they can do that so we've designed a a custom compiler um that allows to to currently compile rust um into a format that our Network can cryptographically understand so you have your application and you define confidential functions and encrypted state that should be processed essed um in a confidential Way by by arum and the way it works really is that on chain on whatever Network you you want to have this confidentiality you have an onchain M Pool where this smart contract then schedules a computation the network picks that up and settles the um publicly auditable result back once once execution has finished so blockchains um make for an an an easily identifiable case um but I'm wondering uh um given the think that you're basically providing off the shelf MPC um what what how what sectors let's say non web3 how would this impact in terms of tangible use cases which increase the use case of privacy in this case yeah that's that's actually a great question because funnily enough um one uh one week ago on on last Monday we um announced our um acquisition of our largest web 2 competitor who've been around for eight years building confidential Computing um for traditional sectors um and more precisely confidential machine learning and confidential AI um also using secure multiparty computation they've done some of the most foundational MPC research um and so they're now part of archium um and um their technology becomes part of arum so over the last two and a half years we've been developing our own MPC protocol and are now combining that with their MPC protocol we haven't really focused on any AI related tasks they have done that um tremendously over the last few years and uh we can combine the trust models and and security models of those two protocols and bring that um into archium so I think um confidential Computing really is one of the big sectors actually that that can bring adoption of of of blockchain systems without whoever um uses confidential Computing having to care about blockchain systems um um I think the use cases that they've tackled is are good examples of that and and MPC really natively um plays nicely with distributed systems having different peers not be be controlled by by one single party I guess um so they've done a lot of um Healthcare um actually also defense intelligence governmental stuff and trfi um as well so I see the uh I see the btb case here I see the compliment portion of your acquisition how does all that translate to cyer punk values so I think um one of the one of the big cyer Punk um value that that we've introduced is um a new kind of sensorship resistance mechanism because um with mpz there's basically different kinds of trust models you can operate in honest majority and in our case dishonest majority so the smallest possible trust trust assumption you don't have to trust that some percentage of of nodes in a multi party computation is honest in our case you only have to trust that there's one single honest participant out of a set of arbitrarily many participants so that's the trust assumption that right from the getg go we said we want to operate in the lowest possible trust assumption so that anyone could could become a participant and have relative Trust of zero because they know they're not maliciously collaborating with all of the other participants right so that's um the trust model that that from the GetGo we we wanted to operate in in order to protect privacy um yeah as as much as possible um the problem you face when when building dishonest majority protocols however from a from a practical standpoint has always been um sort of this high risk um of Dos attacks because while you um only have to trust that there's one arbitrary honest particip ipant that isn't sharing secrets with all of the others one bad participant that wants the computation to fail can make the computation fail um because those those those Legacy protocols basically um didn't allow to pinpoint any cryptographic misbehavior by by such a malicious malicious party so um we've built a protocol that allows to identify that where we can cryptographically identify a misbehaving party that um actively causes the comput ation to fail um because it's something that we don't want we want those confidential computations to be confidential and to be executed when they should be executed um and we can combine that with um yeah game theoretical measures staking and slashing um on on those existing blockchain networks um which which form the sources and then um yeah um call back layers for those confidential computations and so if if those computations fail um someone gets punished um who who cost that that computation to fail so basically adding accountability um and and enforcing confidential execution now I'm interested with arum you're obviously doing well given the acquisition as well and the cases you mentioned as examples um the concept of NPC has been around for um what help me out here since initial research 15 years now conceptually yeah I I I think since since the the ' 80s something like that yeah so so similar to zero knowledge proofs they they had this emergence um starting in the the ' 80s yeah okay um no I think was it FH was in 2009 ex published and that's yeah um so if you look at um the impact especially what MBC could have on the um like I said I mentioned I see the business guys here yeah how would you see this uh impacting and users and their sovereignty um I think mpz actually um has the biggest impact out of um all privacy enhancing technologies that that we've seen zero knowledge proofs and fhe as as the other two cryptographic protocols I feel like mpz will be the most impactful because it allows for more powerful interactions um I feel like um at least why please make it turn yeah so so so at least I I personally have been disillusioned by by CK and in the past we've built a lot of privacy technology with with CK um at arum but um sort of the kind of interactions you can have is are are limited for example operating over shared State encrypted shared state right your knowledge proofs are always I have some secret and I generate a proof about that secret and send it somewhere with MPC that's not required if we if we have this minimum possible zero trust system beat with with MPC or fhe right and we can um just sell some leer to run confidential computations over some encrypted State we don't need um the end user to generate your knowledge proofs to to manage that to do key exchanges with million different users so it has a lot of scalability implications and um implications in the kind of applications we can build you can basically build um a server application right you can have some some server um um I don't know if if we were to live in an Ideal World we could have some trusted server that could just keep all secrets and run those computations right and with with MPC and fhe it's possible to actually do that um to to have such a server where there's no one who can see what's happening we can force that server to actually execute and there's there's no risk of of data leakage so I think that's why um these kind of Technologies are so powerful now um personally comparing MPC with fhe why I feel like MPC is the most powerful um both are using homomorphic encryption we are using homomorphic encryption with with MPC as well um just not the full set of homomorphic operations only a subset and um for for other operations we're we're making use of um yeah so-called pre-processing which just makes it 10,000 times faster so given the the impact you mentioned of scalability uh the fact that you as a user you don't have to run the full compute as yourself what this um could I translate this to that let's say privacy measures would become cheaper for the end user in general is that something you see happening because as of now the trend is um to maintain privacy as an end user um it's becoming more and more expensive just not the stack but also the opportunity cost Etc to actually protect yourself yeah I I I think it's that for one and then it's also composability it's it's not um requiring wallets or or applications to actually do this kind of complex cryptography right it's just communicating with a smart contract and that smart contract then in our case communicates with with our Network right also through a through a smart contract so there's no requirement to to generate those those heavy cryptographic proofs um on on on the client side and um I think the bigger issue really if you look at ccache and and other projects really is less about actually generating proof and rather um just indexing a lot of encrypted State and and Performing cryptographic permutations on that state to to find some some noes so all of that goes away if you're using MPC and operate over encrypted shared state so I think that's why it just enables to do a lot more um important applications also I feel like with the Advent of of AI um there's um yeah just this this important field where where we require um privacy protection um and um yeah the only technology that's not tees um that I've seen really is MPC that can um help with confidential training and inference um in an efficient enough way could you uh describe in this case given the acquisition as well uh with the I looking at the themes of today as well how would these AI data markets look like if they're would be running through not not not even specifically arum but uh given a NPC world yeah so so I I I think what we will be able to facilitate um in the ideal world would be ends to end encrypted AI it would really be having highly sensitive data um be fully encrypted and then using something like archium for collaborative training over that encrypted data so anyone can can have um uh sort of Apple Intelligence on their phone without requiring to trust Apple instead you can you can aggregate the data and have it stored um in an encrypted way and then um have a collaborative um training function be evaluated over all of that distributed encrypted data and then have the output of that training process be an encrypted model um and that's um less um important I think for for privacy and more important for the monetization actually if you want to have this uh closed loop of people willing to provide high quality data because once you have encrypted weights you can have some um smart contract structure in place where people can then run um encrypted inference into those models and pay um the original um data providers who who who who gathered that data used for the training um so I think that will then create a sort of end encrypted AI which can be extremely powerful um I don't know um in in some long-term version having AGI for example if all of the most valuable data is encrypted um you will require those encrypted training processes to actually um train those valuable models yeah well in general before that the compliancy requirements of having a insight into your data set you know for instance Cate issues or other requirement that you and also with gdpr your set needs to be uh private um also by by the way if the audience if you have any questions feel free to jump in of course um happy to um put the fire on our on the guest here um I'm wondering you mentioned in terms of components so we got we fzk you mentioned FH and now with MPC you're providing um um a a a trusted setup for compute what are you missing in the space what could we do better or what are we missing um excellent question I I feel like we're actually doing doing doing pretty well um I think we just need to build a lot more um I mean now with the acquisition we have an amazing team at arum but still um it always takes surprisingly long to to ship things um even once you've done all of the research and and written the papers um implementation takes takes surprisingly long um and um yeah I feel like um I think scalability is still a big concern for for all for all Technologies so um I think all technologies will will greatly benefit from from increased um Hardware acceleration work um I think that's something that we will need um I think in general we've we've seen the trend also with with other teams that that they've realized that um yeah it's it's really not just about proving individual operations locally but rather confidential computations right Global computations to be to be executed in an an encrypted way so I think um I I think there's a trend towards this um which I'm welcoming because I feel like that's that's required um for for um getting more adoption mean like I said um up the supply chain as well uh given the recent nuclear Investments our different data centers to just increase the output basically is that what you're aiming at as we need and uh for the for the advancements and in terms of scalability we need more input at the moment um yeah I think more input will help yeah okay what's holding you back um honest I me I mean um like I ask what what are you missing what's missing space in terms of developments but also what's holding back in terms of if you look at what's um what can you not do even well even though you're building an engine basically for secure Computing yeah I feel like what's what's holding us back um currently um the speed of light um is is holding us back a bit so I I feel like um communication um is still uh just a big constraint for for mpz in general mpz is is the fastest technology but still um communication overhead um be for for pre-processing or um online communication operations um is is a concern um and um yeah we we're doing a lot of research on on that front as well um I think um yeah what's what's holding us back is that it's not 1,000 people working on those problems but I mean we are we are 20 people right now um which is a decently sized team I think but um yeah I think it's just about having more people work on this um in general having more research um having more Engineering in the entire ecosystem um yeah and today you're here at the U ethereum S Congress um as you provide an agnostic Service uh regardless of of type of um vertical or chain even I would say what's your relationship with ethereum so um our our relationship with ethereum I feel like is that um I mean the first Network that we are supporting actually is going to be Solana so um it might be a difficult heard your first um so um yeah we we originated as a as a team from the from the Solana EOS system actually um we were bringing privacy to to Zana um back in the day um we were called elusive um before we uh rebranded to archium um and we were doing onchain privacy um and um so yeah I I I feel like with this with this Evolution to to arum and building generalized confidential Computing we just want to bring this kind of confidential computing power to any network um and um um I feel like ethereum is is incredibly important on that front um secondly I think the relationship to to ethereum is that um it sort of also serves as an inspiration um because I would say in the in this world which is High frut um chains like like Solana or different l2s on eperium um I feel like um yeah what what they've been lacking to a certain degree is also so this um research and and seniority in in cryptographic research so that's what we are bringing to different ecosystems and I feel like um yeah the the work that has been done on on the ethereum layer um by by Cypher punks by by different teams also by by by folks from the cach ecosystem right um serves as sort of an an inspiration I feel like to um to how how to go about things and what values to care about um yeah and how does the as you mentioned you're focusing on Sala first uh the current road map including it scaling through l2s for every ethereum how does it affect you guys as in will your focus be more on them as depti CLI networks or how would you or you to see that so how could you a to the evolution of etheria yes um so so the way I think about it really is that down the line we want to support any kind of network um I think that's it's it's just incredibly important because um yeah I mean what what drives us is to make confidential Computing as fast as possible as cheap as possible and um keep the trustless and censorship resistance at the same time so um do everything at once I guess um but um I feel like that's something that um any network requires um and so I feel like um yeah for us it's just um a question of of logical expansion to to different networks over time and uh in terms of you mention you have a team right now of 20 yeah um ter to advance research in NPC and implementation do you also collaborate on heatons and provide bounties Etc to further research on this or yes asking also for our audience of course so so so basically um this this research um is in-house um most of the research that we are doing so we have a um quite a few of of phds and professors in our team um who've who who've been doing um a lot of mpz and homomorphic encryption research over the years and we're we're um I would say the kind of research that we are doing and looking for is um I would call it more of a practical kind of research um it's it's really looking for for for solutions for efficient practical solutions for different problems that we encounter while we are building out arum um yeah we have time left for one audience question I'm looking at some faces here okay if not then I'll have one um like I said we're here today at the cyer bom Congress there's going to be a lot of parties today um walking around and also talking on stage what would you like to see today what's your expectation or that of arer and of the team um my expectation is uh spreading the word of of mpz um that's what I I like to do um I feel like mpz is still something that people mostly associate with key management for example um and um just spreading awareness of of the power um of of this kind of technology that allows for arbitrary um confidential execution and I feel like should be used in a lot more places thank you so much Yik for today thanks for having me let's give our guest an Applause people [Applause]