CarbyneStackCon '22¶

We’ve come a long way from securely comparing the wealth of two millionaires. We can use MPC to predict satellite collisions, evaluate neural networks, perform cryptographic operations. But I don’t even have an app on my phone to find if the salary range I’m offering overlaps with the expectations of a potential employee. Is the technology here too early? Are researchers working on the wrong problems? Who is MPC really for? This talk is both about barriers and opportunities.

On September 21, 2021, just a little over a year ago, Bosch Research open sourced an internally developed prototype of a cloud platform for Secure Multiparty Computation (MPC) that bridges the gap towards using MPC in an enterprise context by combining state-of-the-art MPC protocols and cloud-native technology. It is time to reflect on what we have achieved so far in terms of further developing the Carbyne Stack platform and building an open source community around it. Knowing the status quo will allow us to shed light on the necessary next steps towards a Carbyne Stack ‘Minimum Viable Platform’ for enterprise use from the Bosch research point of view.

Data-driven insights are critical for modern businesses to inform decision making and optimize processes. In cross-company collaborations necessary data is distributed among companies. Here, data sharing can face trust concerns or privacy issues. But can the data be analyzed without being revealed to anyone? In this talk, we discuss privacy-preserving data collaborations across companies where Carbyne Stack keeps data private and only share insights.

The automotive industry can benefit greatly from the data created on the roads, while at the same time the privacy of the customers is a high aim in this. In this talk, we discuss options to support the use case of privacy-preserving computation in the presence of a scaling number of data providers, analysts, and potential algorithm providers and elaborate on a scalable multi-cloud architecture for the on-demand execution of MPC queries.

Carbyne Stack is based on the SPDZ MPC protocol, its online phase is deployed form the MP-SPDZ library. The necessary cryptographic material, i.e., authenticated Beaver Triples, have to be generated within an offline phase. However, readily available offline phase implementations are prohibitively expensive in terms of computational complexity and bandwidth consumption. Efficient offline phases with low communication can be based on the ring-LPN assumption. Such approaches require to generate sparse shared polynomials, which can be initialized by distributed point functions. This talk is an introduction to the necessary theory and will present the current state of our implementation and the expected cost savings.

In this talk I will present Secrecy, a novel MPC framework for secure relational analytics that we have been developing at Boston University. Secrecy can dramatically improve the performance of complex relational queries while retaining the full security guarantees of the cryptographic protocols. Compared to prior work, Secrecy takes a fundamentally different approach by opening the MPC black boxes to the database optimizer. Secrecy uses an analytical cost model that formulates query costs in terms of secure MPC operations and a query engine for vectorized MPC execution that employs a rich set of logical, physical, and protocol-aware optimizations. These optimizations enable Secrecy to outperform state-of-the-art frameworks and scale to much larger datasets than those reported in prior work. I will describe our ongoing work on integrating Secrecy with the Carbyne stack and conclude with a broader vision on a new generation of systems for privacy-preserving analytics in untrusted clouds.

Some of the most efficient protocols for Multi-Party Computation (MPC) follow a two-phase approach where correlated randomness, in particular Beaver triples, is generated in the offline phase and then used to speed up the online phase. Recently, more complex correlations have been introduced to optimize certain operations even further, e.g. matrix tuples for matrix multiplication, convolution tuples for tensor convolutions and polynomial tuples to evaluate multivariate polynomials. In this talk, we briefly discuss these different forms of generalized tuples, their advantages and trade-offs. Furthermore, we present benchmarks of our corresponding MP-SPDZ-extensions and compare them to state-of-the-art protocols like Overdrive. We show that the use of generalized tuples can significantly improve the performance of MP-SPDZ and of Carbyne Stack.

To be able to run MPC-as-a-Service, we need a platform that performs matchmaking for those wanting to form a compute party, those wanting to secret share their data, and those wanting to perform computations. But who would operate such a platform? This talk will explore how we can utilize web3 technologies to operate this platform in a decentralized way utilizing the latest generation of delegated proof-of-stake blockchains, self-sovereign identities, verifiable credentials, and Carbyne Stack.