libsnark: a C++ library for zkSNARK proofs

---

Authors and contacts

The libsnark library is developed by the SCIPR Lab project and contributors and is released under the MIT License (see the LICENSE file).

Copyright (c) 2012-2017 SCIPR Lab and contributors (see AUTHORS file).

For announcements and discussions, see the libsnark mailing list.

---

[TOC]

---

Overview

This library implements zkSNARK schemes, which are a cryptographic method for proving/verifying, in zero knowledge, the integrity of computations.

A computation can be expressed as an NP statement, in forms such as the following:

• "The C program foo, when executed, returns exit code 0 if given the input bar and some additional input qux."
• "The Boolean circuit foo is satisfiable by some input qux."
• "The arithmetic circuit foo accepts the partial assignment bar, when extended into some full assignment qux."
• "The set of constraints foo is satisfiable by the partial assignment bar, when extended into some full assignment qux."

A prover who knows the witness for the NP statement (i.e., a satisfying input/assignment) can produce a short proof attesting to the truth of the NP statement. This proof can be verified by anyone, and offers the following properties.

• Zero knowledge: the verifier learns nothing from the proof beside the truth of the statement (i.e., the value qux, in the above examples, remains secret).
• Succinctness: the proof is short and easy to verify.
• Non-interactivity: the proof is a string (i.e. it does not require back-and-forth interaction between the prover and the verifier).
• Soundness: the proof is computationally sound (i.e., it is infeasible to fake a proof of a false NP statement). Such a proof system is also called an argument.
• Proof of knowledge: the proof attests not just that the NP statement is true, but also that the prover knows why (e.g., knows a valid qux).

These properties are summarized by the zkSNARK acronym, which stands for Zero-Knowledge Succinct Non-interactive ARgument of Knowledge (though zkSNARKs are also knows as succinct non-interactive computationally-sound zero-knowledge proofs of knowledge). For formal definitions and theoretical discussions about these, see [BCCT12], [BCIOP13], and the references therein.

The libsnark library currently provides a C++ implementation of:

1. General-purpose proof systems:

   1. A preprocessing zkSNARK for the NP-complete language "R1CS" (Rank-1 Constraint Systems), which is a language that is similar to arithmetic circuit satisfiability.

      This zkSNARK construction follows, extends, and optimizes the approach described in [BCTV14a], itself an extension of [BCGTV13], following the approach of [GGPR13] and [BCIOP13]. (An alternative implementation of this approach is the Pinocchio system of [PGHR13].)

      The library also implements a zk-SNARK for R1CS secure in the generic group model [Groth16]. Compared to [BCTV14a] the [Groth16] proof system is faster and achieves shorter proofs, at expense of making stronger security assumptions.

      We provide detailed empirical and asymptotic comparison between these choices.

   2. A preprocessing SNARK for a language of arithmetic circuits, "BACS" (Bilinear Arithmetic Circuit Satisfiability). This simplifies the writing of NP statements when the additional flexibility of R1CS is not needed. Internally, it reduces to R1CS.

   3. A preprocessing SNARK for the language "USCS" (Unitary-Square Constraint Systems). This abstracts and implements the core contribution of [DFGK14]

   4. A preprocessing SNARK for a language of Boolean circuits, "TBCS" (Two-input Boolean Circuit Satisfiability). Internally, it reduces to USCS. This is much more efficient than going through R1CS.

   5. A simulation-extractable preprocessing SNARK for R1CS. This construction uses the approach described in [GM17]. For arithmetic circuits, it is slower than the [BCTV14a] approach, but produces shorter proofs.

   6. ADSNARK, a preprocessing SNARKs for proving statements on authenticated data, as described in [BBFR15].

   7. Proof-Carrying Data (PCD). This uses recursive composition of SNARKs, as explained in [BCCT13] and optimized in [BCTV14b].

2. Gadget libraries (gadgetlib1 and gadgetlib2) for constructing R1CS instances out of modular "gadget" classes.
3. Examples of applications that use the above proof systems to prove statements about:
   1. Several toy examples.
   2. Execution of TinyRAM machine code, as explained in [BCTV14a] and [BCGTV13]. (Such machine code can be obtained, e.g., by compiling from C.) This is easily adapted to any other Random Access Machine that satisfies a simple load-store interface.
   3. A scalable for TinyRAM using Proof-Carrying Data, as explained in [BCTV14b]
   4. Zero-knowldge cluster MapReduce, as explained in [CTV15].

See the above references for discussions of efficiency aspects that arise in practical use of such constructions, as well as security and trust considerations.

This scheme is a preprocessing zkSNARK (ppzkSNARK): before proofs can be created and verified, one needs to first decide on a size/circuit/system representing the NP statements to be proved, and run a generator algorithm to create corresponding public parameters (a long proving key and a short verification key).

Using the library involves the following high-level steps:

1. Express the statements to be proved as an R1CS (or any of the other languages above, such as arithmetic circuits, Boolean circuits, or TinyRAM). This is done by writing C++ code that constructs an R1CS, and linking this code together with libsnark
2. Use libsnark's generator algorithm to create the public parameters for this statement (once and for all).
3. Use libsnark's prover algorithm to create proofs of true statements about the satisfiability of the R1CS.
4. Use libsnark's verifier algorithm to check proofs for alleged statements.

---

The NP-complete language R1CS

The ppzkSNARK supports proving/verifying membership in a specific NP-complete language: R1CS (rank-1 constraint systems). An instance of the language is specified by a set of equations over a prime field F, and each equation looks like: < A, (1,X) > * < B , (1,X) > = < C, (1,X) > where A,B,C are vectors over F, and X is a vector of variables.

In particular, arithmetic (as well as boolean) circuits are easily reducible to this language by converting each gate into a rank-1 constraint. See [BCGTV13] Appendix E (and "System of Rank 1 Quadratic Equations") for more details about this.

---

Elliptic curve choices

The ppzkSNARK can be instantiated with different parameter choices, depending on which elliptic curve is used. The libff library currently provides three options:

• "edwards": an instantiation based on an Edwards curve, providing 80 bits of security.

• "bn128": an instantiation based on a Barreto-Naehrig curve, providing 128 bits of security. The underlying curve implementation is [ate-pairing], which has incorporated our patch that changes the BN curve to one suitable for SNARK applications.

  • This implementation uses dynamically-generated machine code for the curve arithmetic. Some modern systems disallow execution of code on the heap, and will thus block this implementation.

    For example, on Fedora 20 at its default settings, you will get the error zmInit ERR:can't protect when running this code. To solve this, run sudo setsebool -P allow_execheap 1 to allow execution, or use make CURVE=ALT_BN128 instead.

• "alt_bn128": an alternative to "bn128", somewhat slower but avoids dynamic code generation.

Note that bn128 requires an x86-64 CPU while the other curve choices should be architecture-independent; see portability.

---

Gadget libraries

The libsnark library currently provides two libraries for conveniently constructing R1CS instances out of reusable "gadgets". Both libraries provide a way to construct gadgets on other gadgets as well as additional explicit equations. In this way, complex R1CS instances can be built bottom up.

gadgetlib1

This is a low-level library which expose all features of the preprocessing zkSNARK for R1CS. Its design is based on templates (as does the ppzkSNARK code) to efficiently support working on multiple elliptic curves simultaneously. This library is used for most of the constraint-building in libsnark, both internal (reductions and Proof-Carrying Data) and examples applications.

gadgetlib2

This is an alternative library for constructing systems of polynomial equations and, in particular, also R1CS instances. It is better documented and easier to use than gadgetlib1, and its interface does not use templates. However, fewer useful gadgets are provided.

---

Security

The theoretical security of the underlying mathematical constructions, and the requisite assumptions, are analyzed in detailed in the aforementioned research papers.

** This code is a research-quality proof of concept, and has not yet undergone extensive review or testing. It is thus not suitable, as is, for use in critical or production systems. **

Known issues include the following:

• The ppzkSNARK's generator and prover exhibit data-dependent running times and memory usage. These form timing and cache-contention side channels, which may be an issue in some applications.

• Randomness is retrieved from /dev/urandom, but this should be changed to a carefully considered (depending on system and threat model) external, high-quality randomness source when creating long-term proving/verification keys.

---

Build instructions

Dependencies

The libsnark library relies on the following:

• C++ build environment
• CMake build infrastructure
• GMP for certain bit-integer arithmetic
• libprocps for reporting memory usage
• Fetched and compiled via Git submodules:
  • libff for finite fields and elliptic curves
  • libfqfft for fast polynomial evaluation and interpolation in various finite domains
  • Google Test (GTest) for unit tests
  • ate-pairing for the BN128 elliptic curve
  • xbyak just-in-time assembler, for the BN128 elliptic curve
  • Subset of SUPERCOP for crypto primitives needed by ADSNARK

So far we have tested these only on Linux, though we have been able to make the libsnark work, with some features disabled (such as memory profiling or GTest tests), on Windows via Cygwin and on Mac OS X. See also the notes on portability below. (If you port libsnark to additional platforms, please let us know!)

Concretely, here are the requisite packages in some Linux distributions:

• On Ubuntu 16.04 LTS:

    $ sudo apt-get install build-essential cmake git libgmp3-dev libprocps4-dev python-markdown libboost-all-dev libssl-dev
  

• On Ubuntu 14.04 LTS:

    $ sudo apt-get install build-essential cmake git libgmp3-dev libprocps3-dev python-markdown libboost-all-dev libssl-dev
  

• On Fedora 21 through 23:

    $ sudo yum install gcc-c++ cmake make git gmp-devel procps-ng-devel python2-markdown
  

• On Fedora 20:

    $ sudo yum install gcc-c++ cmake make git gmp-devel procps-ng-devel python-markdown
  

Building

Fetch dependencies from their GitHub repos:

$ git submodule init && git submodule update

Create the Makefile:

$ mkdir build && cd build && cmake ..

Then, to compile the library, tests, and profiling harness, run this within the build directory:

$ make

To create the HTML documentation, run

$ make doc

and then view the resulting README.html (which contains the very text you are reading now).

To compile and run the tests for this library, run:

$ make check

Using libsnark as a library

To develop an application that uses libsnark, it's recommended to use your own build system that incorporates libsnark and dependencies. If you're using CMake, add libsnark as a git submodule, and then add it as a subdirectory. Then, add snark as a library dependency to the appropriate rules.

To build and install the libsnark library:

$ DESTDIR=/install/path make install

This will install libsnark.a into /install/path/lib; so your application should be linked using -L/install/path/lib -lsnark. It also installs the requisite headers into /install/path/include; so your application should be compiled using -I/install/path/include.

In addition, unless you use WITH_SUPERCOP=OFF, libsnark_adsnark.a will be installed and should be linked in using -lsnark_adsnark.

When you use compile your application against libsnark, you must have the same conditional defines (#define FOO or g++ -DFOO) as when you compiled libsnark, due to the use of templates. One way to figure out the correct conditional defines is to look at build/libsnark/CMakeFiles/snark.dir/flags.make after running cmake. (Issue #21)

Building on Windows using Cygwin

Install Cygwin using the graphical installer, including the g++, libgmp, cmake, and git packages. Then disable the dependencies not easily supported under CygWin, using:

$ cmake -DWITH_PROCPS=OFF ..

Building on Mac OS X

On Mac OS X, install GMP from MacPorts (port install gmp). Then disable the dependencies not easily supported under OS X, using:

$ cmake -DWITH_PROCPS=OFF ..

MacPorts does not write its libraries into standard system folders, so you might need to explicitly provide the paths to the header files and libraries by appending CXXFLAGS=-I/opt/local/include LDFLAGS=-L/opt/local/lib to the line above.

---

Build options

The following flags change the behavior of the compiled code. Use

 $ cmake -Dname1=ON -Dname2=OFF ...

to control these (you can see the default at the top of CMakeLists.txt).

• cmake -DCURVE=choice (where choice is one of: ALT_BN128, BN128, EDWARDS, MNT4, MNT6)

  Set the default curve to one of the above (see elliptic curve choices).

• cmake -DLOWMEM=ON

  Limit the size of multi-exponentiation tables, for low-memory platforms.

• cmake -DWITH_PROCPS=OFF

  Do not link against libprocps. This disables memory profiling.

• cmake -DWITH_SUPERCOP=OFF

  Do not link against SUPERCOP for optimized crypto. The ADSNARK executables will not be built.

• cmake -DMULTICORE=ON

  Enable parallelized execution of the ppzkSNARK generator and prover, using OpenMP. This will utilize all cores on the CPU for heavyweight parallelizable operations such as FFT and multiexponentiation. The default is single-core.

  To override the maximum number of cores used, set the environment variable OMP_NUM_THREADS at runtime (not compile time), e.g., OMP_NUM_THREADS=8 test_r1cs_sp_ppzkpc. It defaults to the autodetected number of cores, but on some devices, dynamic core management confused OpenMP's autodetection, so setting OMP_NUM_THREADS is necessary for full utilization.

• cmake -DUSE_PT_COMPRESSION=OFF

  Do not use point compression. This gives much faster serialization times, at the expense of ~2x larger sizes for serialized keys and proofs.

• cmake -DMONTGOMERY_OUTPUT=ON (enabled by default)

  Serialize Fp elements as their Montgomery representations. If this option is disabled then Fp elements are serialized as their equivalence classes, which is slower but produces human-readable output.

• cmake -DBINARY_OUTPUT=ON (enabled by default)

  In serialization, output raw binary data (instead of decimal), which is smaller and faster.

• cmake -DPROFILE_OP_COUNTS=ON

  Collect counts for field and curve operations inside static variables of the corresponding algebraic objects. This option works for all curves except bn128.

• cmake -DUSE_ASM=ON (enabled by default)

Use architecture-specific assembly routines for F[p] arithmetic and heap in
multi-exponentiation. (If disabled, use GMP's `mpn_*` routines instead.)

• cmake -DUSE_MIXED_ADDITION=ON

  Convert each element of the proving key and verification key to affine coordinates. This allows using mixed addition formulas in multiexponentiation and results in slightly faster prover and verifier runtime at expense of increased generator runtime.

• cmake -DPERFORMANCE=ON

  Enables compiler optimizations such as link-time optimization, and disables debugging aids. (On some distributions this causes a plugin needed to handle lto object link error and undefined references, which can be remedied by AR=gcc-ar make ....)

• cmake -DOPT_FLAGS=...

  Set the C++ compiler optimization flags, overriding the default (e.g., -DOPT_FLAGS="-Os -march=i386").

• cmake -DDEPENDS_DIR=...

  Sets the dependency installation directory to the provided absolute path (default: installs dependencies in the respective submodule directories)

• cmake -DUSE_LINKED_LIBRARIES=ON

  Setting this flag enables CMake to include installed libfqfft and libff libraries. This will tell the compiler to ignore the libfqfft and libff dependencies provided in the depends folder.

Not all combinations are tested together or supported by every part of the codebase.

---

Docker

You can run libsnark on Docker:

$ docker build -t libsnark .
$ docker run -ti libsnark /bin/bash

---

Tutorials

libsnark includes a tutorial, and some usage examples, for the high-level API.

• libsnark/gadgetlib1/examples1 contains a simple example for constructing a constraint system using gadgetlib1.

• libsnark/gadgetlib2/examples contains a tutorial for using gadgetlib2 to express NP statements as constraint systems. It introduces basic terminology, design overview, and recommended programming style. It also shows how to invoke ppzkSNARKs on such constraint systems. The main file, tutorial.cpp, builds into a standalone executable.

• libsnark/zk_proof_systems/ppzksnark/r1cs_ppzksnark/profiling/profile_r1cs_ppzksnark.cpp constructs a simple constraint system and runs the ppzksnark. See below for how to run it.

---

Executing profiling example

The command

 $ libsnark/zk_proof_systems/ppzksnark/r1cs_ppzksnark/profiling/profile_r1cs_ppzksnark 1000 10 Fr

exercises the ppzkSNARK (first generator, then prover, then verifier) on an R1CS instance with 1000 equations and an input consisting of 10 field elements.

(If you get the error zmInit ERR:can't protect, see the discussion above.)

The command

 $ libsnark/zk_proof_systems/ppzksnark/r1cs_ppzksnark/profiling/profile_r1cs_ppzksnark 1000 10 bytes

does the same but now the input consists of 10 bytes.

---

Portability

libsnark is written in fairly standard C++11.

However, having been developed on Linux on x86-64 CPUs, libsnark has some limitations with respect to portability. Specifically:

1. libsnark's algebraic data structures assume little-endian byte order.

2. Profiling routines use clock_gettime and readproc calls, which are Linux-specific.

3. Random-number generation is done by reading from /dev/urandom, which is specific to Unix-like systems.

4. libsnark binary serialization routines (see BINARY_OUTPUT above) assume a fixed machine word size (i.e. sizeof(mp_limb_t) for GMP's limb data type). Objects serialized in binary on a 64-bit system cannot be de-serialized on a 32-bit system, and vice versa. (The decimal serialization routines have no such limitation.)

5. libsnark requires a C++ compiler with good C++11 support. It has been tested with g++ 4.7 and newer, and clang 3.4 and newer.

6. On x86-64, we by default use highly optimized assembly implementations for some operations (see USE_ASM above). On other architectures we fall back to a portable C++ implementation, which is slower.

7. The ate-pairing library, require by the BN128 curve, can be compiled only on i686 and x86-64. (On other platforms, use other -DCURVE=... choices.)

8. The SUPERCOP library, required by ADSNARK, can be compiled only on i686 and x86-64. (On other platforms, use -DWITH_SUPERCOP=OFF.)

Tested configurations include:

• Debian jessie with g++ 4.7 on x86-64
• Debian jessie with clang 3.4 on x86-64
• Fedora 20/21 with g++ 4.8.2/4.9.2 on x86-64
• Fedora 21 with g++ 4.9.2 on x86-32, for non-BN128 curves (-DWITH_SUPERCOP=OFF)
• Ubuntu 14.04 LTS with g++ 4.8 on x86-64
• Ubuntu 14.04 LTS with g++ 4.8 on x86-32, for non-BN128 curves (-DWITH_SUPERCOP=OFF)
• Ubuntu 15.04/16.04 LTS with g++ 4.9.2/5.3.1 on ARM AArch32/AArch64, for non-BN128 curve choices
• Debian wheezy with g++ 4.7 on ARM little endian (Debian armel port) inside QEMU, for EDWARDS and ALT_BN128 curve choices
• Windows 7 with g++ 4.8.3 under Cygwin 1.7.30 on x86-64 for EDWARDS and ALT_BN128 curve choices (-DWITH_PROCPS=OFF and GTestdisabled)
• Mac OS X 10.9.4 (Mavericks) with Apple LLVM version 5.1 (based on LLVM 3.4svn) on x86-64 (-DWITH_PROCPS=OFF and GTest disabled)

---

Directory structure

The directory structure of the libsnark library is as follows:

• libsnark: main C++ source code, containing the following modules:
  • common: miscellaneous utilities
  • gadgetlib1: gadgetlib1, a library to construct R1CS instances
    • gadgets: basic gadgets for gadgetlib1
  • gadgetlib2: gadgetlib2, a library to construct R1CS instances
  • relations: interfaces for expressing statement (relations between instances and witnesses) as various NP-complete languages
    • constraint_satisfaction_problems: R1CS and USCS languages
    • circuit_satisfaction_problems: Boolean and arithmetic circuit satisfiability languages
    • ram_computations: RAM computation languages
  • zk_proof_systems: interfaces and implementations of the proof systems
  • reductions: reductions between languages (used internally, but contains many examples of building constraints)
• depends: external dependencies which are automatically fetched and compiled (overridable by cmake -DDEPENDS_DIR=...)

Some of these module directories have the following subdirectories:

• ...
  • examples: example code and tutorials for this module
  • tests: unit tests for this module

In particular, the top-level API examples are at libsnark/r1cs_ppzksnark/examples/ and libsnark/gadgetlib2/examples/.

---

Further considerations

Multiexponentiation window size

The ppzkSNARK's generator has to solve a fixed-base multi-exponentiation problem. We use a window-based method in which the optimal window size depends on the size of the multiexponentiation instance and the platform.

On our benchmarking platform (a 3.40 GHz Intel Core i7-4770 CPU), we have computed for each curve optimal windows, provided as fixed_base_exp_window_table initialization sequences, for each curve; see X_init.cpp for X=edwards,bn128,alt_bn128.

Performance on other platforms may not be optimal (but probably not be far off). Future releases of the libsnark library will include a tool that generates optimal window sizes.

---

References

[BBFR15] ADSNARK: nearly practical and privacy-preserving proofs on authenticated data , Michael Backes, Manuel Barbosa, Dario Fiore, Raphael M. Reischuk, IEEE Symposium on Security and Privacy (Oakland) 2015

[BCCT12] From extractable collision resistance to succinct non-Interactive arguments of knowledge, and back again , Nir Bitansky, Ran Canetti, Alessandro Chiesa, Eran Tromer, Innovations in Computer Science (ITCS) 2012

[BCCT13] Recursive composition and bootstrapping for SNARKs and proof-carrying data Nir Bitansky, Ran Canetti, Alessandro Chiesa, Eran Tromer, Symposium on Theory of Computing (STOC) 13

[BCGTV13] SNARKs for C: Verifying Program Executions Succinctly and in Zero Knowledge , Eli Ben-Sasson, Alessandro Chiesa, Daniel Genkin, Eran Tromer, Madars Virza, CRYPTO 2013

[BCIOP13] Succinct non-interactive arguments via linear interactive Proofs , Nir Bitansky, Alessandro Chiesa, Yuval Ishai, Rafail Ostrovsky, Omer Paneth, Theory of Cryptography Conference 2013

[BCTV14a] Succinct non-interactive zero knowledge for a von Neumann architecture , Eli Ben-Sasson, Alessandro Chiesa, Eran Tromer, Madars Virza, USENIX Security 2014

[BCTV14b] Scalable succinct non-interactive arguments via cycles of elliptic curves , Eli Ben-Sasson, Alessandro Chiesa, Eran Tromer, Madars Virza, CRYPTO 2014

[CTV15] Cluster computing in zero knowledge , Alessandro Chiesa, Eran Tromer, Madars Virza, Eurocrypt 2015

[DFGK14] Square span programs with applications to succinct NIZK arguments , George Danezis, Cedric Fournet, Jens Groth, Markulf Kohlweiss, ASIACCS 2014

[Groth16] On the Size of Pairing-based Non-interactive Arguments , Jens Groth, EUROCRYPT 2016

[GM17] Snarky Signatures: Minimal Signatures of Knowledge from Simulation-Extractable SNARKs , Jens Groth and Mary Maller, IACR-CRYPTO-2017

[GGPR13] Quadratic span programs and succinct NIZKs without PCPs , Rosario Gennaro, Craig Gentry, Bryan Parno, Mariana Raykova, EUROCRYPT 2013

[ate-pairing] High-Speed Software Implementation of the Optimal Ate Pairing over Barreto-Naehrig Curves , MITSUNARI Shigeo, TERUYA Tadanori

[PGHR13] Pinocchio: Nearly Practical Verifiable Computation , Bryan Parno, Craig Gentry, Jon Howell, Mariana Raykova, IEEE Symposium on Security and Privacy (Oakland) 2013