https://twitter.com/kanzure/status/1090741190100545536

Hi. I am Alessandro Chiesa, a faculty member at UC Berkeley. I work on cryptography. I'm also a chief scientist at Starkware. I was a founder of zcash. Today I will like to tell you about recent work in cryptographic proof systems. This is joint work with people at Starkware, UC Berkeley, MIT Media Lab, and others. In the prior talk, we talked about scalability. But this current talk is more about privacy by using cryptographic proofs.

paper: https://eprint.iacr.org/2018/828

Let me spend just one slide to remind you what zkSNARKs are. They are a cryptographic proof system that allows you to produce non-interactive proofs for computational integrity. You can say I know w such that y = F(w) without revealing w. The privacy property here is called zero-knowledge. The proof itself provably hides the secret. The efficient properties of SNARKs have to do with the size of the proof. Succinctness dictates that the size of the proof only grows polylogarithmically in the length of the proof being produced. In the previous talk, we also focused on succinctness of the verifier. The process of verifying the proof must be exponentially faster than making a new proof.

Over the past few years, there have been many beautiful and efficient constructions of zero-knowledge proofs. We've seen real world deployments like zcash and so on.

We wanted to study a class of zkSNARKs that are transparent. These are those that don't rely on any ... any public parameters are just pure public parameters. I am not going to give an overview of trusted setup, I am sure there's enough resources online about trusted setup.

Most instructions have two parts. You can think about a frontend, a reduction that takes a problem you care about, and a statement about the computation, and reduce it to some intermediate representation that involves statements. You have a proof system that picks up the intermediate representation, and produce a proof string that represents some properties about the intermediate representation. The choice of intermediate representation is very important because it impacts efficiency of reduction and the efficiency of the proof system.

Today I will tell you about a proof system for an intermediate representation that has shown strong properties.

How many of you have heard of R1CS before? Okay, so about a third of you. That's great. R1CS construction is a very simple language. Let me define it for you in a little box. It says let give you three things, A, B, and C and some public input x. The problem you're trying to solve is, does there exist an input w such that if you take A and multiply it by the vector xw, and take B and multiply by vector xw, and C and do the same, you get three vectors that are.... You use an element-wise product, where you multiply by coordinates. So does the first coordinate of the first vector multiplied by the second, equal the first coordinator of the last vector?

Why are we looking at this funny looking problem? It generalizes a rather natural type of computation which is circuit computations. Circuits might not sound natural, but they are natural for cryptographic proof systems. It's a type of problem from which you can construct proof systems. R1CS is a more general problem within that. It's easy to make certain circuits in this system. It's a simple linear algebraic structure which simplifies the backend. It generalizes arithmetic circuit SAT on the frontend. It's a good tradeoff between frontend and backend.

This is not just in principle, but we have a lot of empirical evidence that real world programmers out there have been able to express problems of interest in this language. They built zcash, libsnark, xJsnark, ZoKrates, Snarky, etc.

In this work, we put forward a construction of a zkSNARK that does not have a trusted setup. It only relies on public randomness and cryptographic hash functions. It lets you prove satisfiability of rank-1 satisfaction constraint problems. The parameters of this proof system are a good verifying itme, and we get very small proofs- they are exponentially smaller than the computation. The number of gates or constraints in your circuit, the proof size only grows polylogarithmically in the size of your computation. That's the key efficiency property we're trying to achieve here.

In terms of real world numbers, the proving time for millions of gates is mere minutes. The size of this proof is going to be on the order of 10-20 of kilobytes. Verification will be in a few seconds for millions of gates.

Why are we studying this type of construction? Well, they have this transparent setup. It has a black box use of symmetric-key crypto (a random oracle), and it's plausibly post-quantum secure and we have very few constructions of SNARKs that can withstand quantum adversaries.

The first proof size achieved in Aurora is 50x larger than the proofs in bulletproofs. It's another proof systme for circuits that has a transparent setup. These proof systems are not competitive with other constructions that rely on public-key cryptography that relies on the hardness of discrete logs.

On the other hand, what we achieve in Aurora, is that among constructions that use ... we achieve 10x improvement in the proof size compared to prior works that when applied ot these circuits. So, the punchline is that in this work we achieved the smallest known-to-date post-quantum SNARK for circuits of the R1CS problem.

I'd like to tell you about what's happening in these SNARKs and where does this class of post-quantum SNARKs come from. The first SNARK discovered was already done in the cryptography community in the mid 90s. Back then we didn't call them SNARKs, they had another name. We had constructions where if you give me a probabilistically checkable proof- which you can check by querying a few locations in the proof- then using such a proof you can build a SNARK. Already back then, this construction had a property that today we recognize is very attractive such as a transparent setup, you only use random oracles and cryptographic hash functions, and it was already post-quantum secure plausibly back in the 90s. But PCPs were treated as bad constructions because they were making use of strong cryptographic assumptions (namely the random oracle) so for the most part they were not treated as a satisfying construction.

From a practical perspective, for us, these are great properties. They are easy to deploy, they are plausibly post-quantum secure, and they don't need new crypto. So why not deploy these? Well, we still don't have good asymptotic results for PCPs or even good implementations of this object. Okay, well, it means that whatever I'm discussing today is not built from PCPs.

So how do we achieve post-quantum SNARKs today?

A few years ago, in a paper, we proposed an alternative approach where we extended PCPs with interactive oracle proofs. It's a multi-round analog proof, and over the past few years we have been able to show that in this model we can still get all the previous advantages of PCPs while circumventing many of the issues. We were able to achieve interactive oracle proofs with optimal proof size. We were also able to put together complete efficiency prototypes with..... The STARKs in the previous talk had an interactive oracle proof underneath. The scalability properties come from advances and limtiations of interactive oracle proofs on which we layer other crypto.

If we want to focus on the language of R1CS, then it sounds reasonable that -- it shouldn't be surprising that the core of the work is designing an interactive oracle proof for this problem. I want to design such a proof for checking satisfiyability of this problem of rank-one constraint satisfaction problem. We are able to show that for this problem you are able to achieve opitmal proof size, this is the size of the proof you're checking which is linear. We can generate small proofs. Are people still following? Something I like about this work is that the construction behind this work isn't that difficult. If I had a full hour, I could actually give you a reasonable picture of what's happening in the protocol itself and you'd have some sense for what's happening.

Instead, I will tell you a little bit about the high-level structure of how you go about taking a problem like R1CS and checking it with small proofs. The problem is, I give you three matrices, a public vector x, and I want to ask if I can extend this public vector with a secret extension such that a certain constraint is kept. This looks like a circuit.

The naive way of checking this problem would be for the prover to send to the verifier the secret value. I could just send you z, which is large, almost as large as the computation so I haven't saved everything. Okay, fine. Let's make the problem simpler. What about instead of r1cs, we can do a sub-problem that can communicate less information. Instead of sending z, I send A, B, C and z, and then you as a verifier check that A, B, and C are the correct linear transformations of z and that A, B, and C given that they are correct linear transformations, then they satisfy a coordinate-wise product. Whatever I was checking before, I have translated it into four subproblems that are equivalent to what I was checking before.

In the prior talk, at some point the word "codes" was mentioned. In all of these STARK constructions, error correcting codes play an important role because you take all of these computation traces (like check signatures, check paths, etc.). What we do with this computation trace is, what happens in all these proof systems including SNARKs in zcash, we take those computation traces and encode them using special codes that give us some error correcting properties that later give rise to the properties of probabilistic checking.

In these constructions, the code of choice is the Reed-Solomon code. In the theory of error correcting codes, this is a foundational code. When you want to encode information, you're going to pick some field that is large enough. And you pick two sets- one is called the interpolation set. That's called H. Inside the interpolation set, you lay out your computation trace. Then you pick an evaluation set. What do you do with these sets? First you think about, that I can put a computation trace inside the interpolation set, and then you interpolate. That's why we call it an interpolation set. You get a unique polyonomial that includes information inside it about the trace. This evaluation has information inside the computation trace, but it has it in some sort of redundant form. Essentially, two different computation traces even if they differ at one step, now in only one step, wants to encode them, they will differ in many many locations. So it amplifies differences between different execution traces.

Intuitively, this should have something t odo with catching errors in a computation with small number of queries. If small differences create big differences in encoding, then maybe the prover would have a harder time cheating. But this is only at a high level. This is just to suggest why there are many roles for codes to play here.

At the very high level, we translate each of these sub-problems that we have to check into corresponding questions about polynomials. Instead of talking about vectors and matrices, now we talk about operations on polynomials.

The way that our interactive oracle proof works in our work works, we have subprotocols for these linear relations (lincheck) and for checking the coordinate-wise product we call it rowcheck. And lincheck reduces to univariate sumcheck, and there's Reed-Solomon testing as well. For rowcheck there's standard PCP checks.

Enough with technical things. We've been developing a library called libiop. It enables the construction of post-quantum SNARKs starting with interactive oracle proofs. It's a C++ toolchain that we hope to put online in the next couple months. This library enables you to make constructions of zkSNARKs with transparent setup (only random oracle model), lightweight crypto (only random oracle model), and it's post-quantum secure. There are also other components, we automize the construction of a post-quantum SNARK and we show how to construct not just Aurora but other protocols from the literature such as Ligero or direct LDT or FRI LDT.

I want to thank Dev Ojha at UC Berkeley who has been doing lots of great work on improving and cleaning up this library for an upcoming open-source release.

The key property that we have achieved in this work is the shortest proof size of circuits. It grows log squared in the size of the circuit. I hope that you have learned something about post-quantum SNARKs today, and I'd be happy to answer questions.