Skip to content

feat(circuits): prove synchronized threshold decryption in C6/C7 #2050

Open

Description

@neogiacomo

Parent: #2456 · Depends on: #2458, #2459

Delivery alignment (2026-11-09)

Context

fhe.rs v0.6.0 provides trbfv partial decryption to the synchronized-decryptor construction in Colin de Verdière–Passelègue–Stehlé, ePrint 2026/031, §5. For a designated set S of threshold + 1 parties, with 1-based IDs, a party now produces

d_i^(S,ct) = λ_i^S · c1 · sh_i + e_i + r_i^(S,ct), where r_i = Σ_{j∈S,j≠i}(F_{k_i,j}(S,ct) − F_{k_j,i}(S,ct)).

Each party samples its own fresh smudging noise e_i; the masks cancel only for a complete set of shares for the same S and ct. Final decryption uses c0 + Σ_{i∈S} d_i, without applying Lagrange weights again. v0.6.0 implements the PRF using Poseidon2/SAFE; its PRF encoding and key layout must be the source of truth for circuit interoperability.

Interfold's current C6 instead proves d_i = c0 + c1·sh_i + e_sm,i using the C4 aggregated noise-share commitment. C7 then Lagrange-interpolates these shares. C1/C2b/C4b and the Rust witness/proof wiring still support shared smudging noise. Merely upgrading fhe.rs would leave the proofs checking a different decryption algorithm.

Circuit-side work

  • Define a versioned proof/public-input contract for the synchronized mode. Bind each C6 share to the same designated set S (canonical ordered IDs), ciphertext/level, E3 domain and party ID. Bind C7 to that exact context and to the corresponding C6 outputs; reject duplicate, out-of-range, unauthorized, missing or mixed-set/ciphertext shares. Preserve the C6→C7 share-commitment link.
  • Update C6's proven relation to λ_i^S·c1·sh_i + e_i + r_i^(S,ct); keep the C4 secret-key commitment link, but do not accept an arbitrary uncommitted mask/noise witness. Constrain the local smudging noise to its signed integer bound and canonical RNS encoding, with parameter-compatible total noise. A circuit cannot prove randomness/freshness of local sampling alone: specify the host-side one-time/retry policy separately rather than claiming that range checks do so.
  • Specify and implement private PRF-key provenance and mask verification: authenticate each party's incoming/outgoing keys as the correctly paired committee keys established during setup, without exposing them; constrain Poseidon2/SAFE digest/evaluation, domain separation, key-to-field mapping, canonical S, ciphertext coefficient order/level, output-to-CRT-residue reduction, and signed pairwise sums to match v0.6.0 byte/field-for-field. If setup currently lacks a sound key-binding mechanism, treat it as a prerequisite, not as an unconstrained C6 input. Assess constraint/proving costs before fixing the final design (the PRF emits a value per ciphertext coefficient and CRT limb).
  • Update C7 to sum exactly the committed C6 shares plus c0 once, then perform the existing CRT reconstruction and plaintext decoding; remove C7's Lagrange interpolation. Retire/version the C1→C2b→C4b shared-smudging commitment path for this mode without breaking in-flight E3s or historical proof formats.
  • Update Noir entry points, Rust witness generation (crates/zk/witness and crates/zk/config), prover/commitment links and recursive/public-input layouts, parameter/codegen and configuration manifests, verifier keys/deployments and docs together. Wire proof requests to the same S, ciphertext and PRF key material used by the host-side fhe.rs implementation. Do not reuse old C6/C7 VKs or silently mix protocol versions.

Acceptance criteria

  • A valid end-to-end C6→C7 proof decrypts with v0.6.0-compatible witnesses on supported parameter/committee profiles. Initial integration uses one ciphertext and a fixed T+1 designated set; proposed selection is the lowest canonical T+1 IDs from the published H-roster, subject to protocol review. Circuit/PRF conformance vectors should still cover differing ciphertexts and canonical non-prefix sets; those vectors do not enable multiple outputs or runtime set replacement.
  • Negative proofs fail for wrong/mixed S, ciphertext or level, party/key pairing, tampered PRF transcript/output, out-of-bound smudging, arbitrary d_i, duplicate/missing shares and inconsistent C6→C7 commitments. Negative circuit tests ensure incomplete shares across sets cannot form a valid C7 proof. Runtime retries with a replacement S on the same ciphertext require the separate reviewed design in Design synchronized designated-set replacement and recovery #2462; multiple authorized decryptions per E3 require Design multiple synchronized ciphertext decryptions per E3 #2463.
  • Public-input/commitment links and generated VKs match the versioned circuit artifacts; measure gates/proving time for representative insecure and secure profiles and obtain a cryptographic review of the key-binding and PRF-in-circuit design before production rollout.

Related but distinct: C6 optimization #1947 concerns the existing circuit, not this protocol migration.

Activity

  1. neogiacomo commented on Nov 19, 2026

    @neogiacomo
    ContributorAuthor

    Review handoff: disabled Minimum decryption is now qualified at both degrees. Secure C6/C7/recursive were 8,609,254 / 1,516,802 / 2,242,265 gates; 26.68 GiB sampled peak with a 32 GiB stop guard. Native plaintext, both child VK pins, the outer bb proof, and generated Solidity verification were checked. Secure setup inputs were native; the signed DKG + actual registry + request-aware wrapper path was exercised on insecure Minimum. No secure registry route, deployment or runtime admission is claimed. Candidate manifests, reproduction and timings: #2530.

    Merge order after independent approval: #2529 (CI selects and verifies an archive matching the checkout SS catalogue), #2524 (PRF), #2525 (C6), #2526 (C7), #2527 (recursive), #2528 (registry/context wrapper), #2530 (secure + manifests). Retarget child PRs to main in sequence; do not merge children into parent branches as a substitute for review. Each diff is below 100 files.

    @YounesTal1: independent crypto/protocol review requested on the key pairing/openings, native PRF transcript, integer noise and quotient bounds, secure product evaluation challenge and exact recursive/request binding. This is a request for review, not a sign-off. #2460 retains ownership of live fixed-set retries/failover and host freshness; #2462/#2463 remain separate.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Labels

cryptographyConcerned with cryptographysecurityRelevant to security

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions