swisspost-evoting-go-poc/ARCHITECTURE.md

Multi-Party Architecture

This PoC emulates the Swiss Post e-voting system's trust structure: the parties that in production run on separate machines under separate operators are modeled here as separate in-process endpoints that communicate only through an authenticated, confidential transport. No party reaches into another's private state; every inter-party message is Ed25519-signed and verified, and confidential channels are keyed by X25519 ECDH — all of that cryptography implemented in Rust (pkg/transportsec → rust/transportsec).

Scalability is explicitly out of scope (it remains a single-binary PoC). What is in scope is full cryptographic emulation of the trust boundaries.

Parties

Party	Role	Private state (never leaves the party)
Setup component (SDM analog)	Generates election parameters, assembles voting cards and the return-codes mapping table	election-event seed, assembled card secrets
CC0–CC3 (4 control components)	Split-trust key generation; ballot proof checks; mix-net shuffle + partial decryption	each CC's ElGamal secret key, return-code secret, shuffle permutation & randomness
Electoral Board (offline)	Final shuffle and decryption on the air-gapped side	EB secret key, board passwords
Voting server / ballot box	Receives ballots, drives return-code extraction, stores the ballot box	mapping table, ballot box contents
Voter client	Encrypts the ballot, produces ballot proofs, checks return codes	card secrets, encryption randomness, vcSK
Verifier / auditor	Independently re-checks the entire public transcript	nothing — public data only

Transport security (all crypto in Rust)

             ┌──────────────── Ed25519 root CA (Rust-signed X.509) ───────────────┐
             │                                                                     │
      issues identity certs (Ed25519, no RSA) binding party name → Ed25519 pubkey  │
             │                                                                     │
   ┌─────────▼──────────┐   signed Envelope    ┌──────────▼─────────┐
   │  Party A (Identity)│ ───────────────────▶ │  Party B (Identity)│
   │  Ed25519 seed      │  Ed25519 sig (Rust)  │  verifies sig (Rust)│
   │  X25519 keypair    │ ◀─────────────────── │                    │
   └────────────────────┘   secure channel     └────────────────────┘
                         X25519 ECDH (Rust) → AES-256-GCM payloads

• Identity (pkg/transport/identity.go): each party holds an Ed25519 signing key, an X25519 ECDH key, and an X.509 certificate. The CA signs certificates through a crypto.Signer shim whose Sign calls the Rust Ed25519 — so even x509.CreateCertificate's signature bytes come from Rust. Certificate verification extracts the TBS bytes and calls the Rust verifier (never Go's x509 internals).
• Envelope (envelope.go): a signed message {from,to,type,nonce,payload}. The signature covers an injective length-prefixed encoding of all fields.
• SecureChannel (channel.go): X25519 ECDH shared secret → session key → AES-256-GCM. Used where a payload must be confidential, not just authentic.
• Directory + Bus (bus.go): the directory admits a party only after its cert chains to the CA, binding the name to the cert's Ed25519 key. The bus routes envelopes and verifies every message's signature (request and reply) before delivery — this is where authenticity is enforced at the boundary.

Why EdDSA / ECDH instead of the production RSA

The production system uses RSASSA-PSS signatures and RSA-based channel security. This PoC deliberately substitutes an elliptic-curve stack — Ed25519 for signatures, X25519 for key agreement — with no RSA anywhere, including the X.509 CA. This is a design choice for the PoC, not a claim about the production system.

Message flow (implemented)

Run it with evote netdemo --verbose to watch every signed envelope.

• Setup — setup → CCj: gen-cc-keys; CCj → setup: cc-keys (PK + Schnorr proofs, verified on receipt); setup → EB: gen-eb-key; EB → setup: eb-key. Setup combines the keys and publishes them to the transcript.
• Cards — for each voter, setup → CCj: long-code-share-req; CCj → setup: long-code-share-resp (return-code shares from the CC's private secret). Setup assembles cards and delivers them confidentially: setup → voter: voting-card and setup → server: mapping-table (both X25519-encrypted).
• Voting — voter → server: cast-ballot (ballot + E2 + equality proof); server → CCj: verify-ballot; CCj → server: ballot-verdict. Stored on unanimous acceptance (vcPK persisted). Then the return-code extraction: server → CCj: rc-exp-req / rc-dec-req; CCj →: rc-exp-resp / rc-dec-resp. The server returns the recovered code to the voter, who checks it against the card (cast-as-intended — see below).
• Tally — → server: start-tally (server pads the ballot box); then → CCj: shuffle / CCj →: shuffled chained through all CCs; → EB: final-mix / EB →: final-done. Each party posts its shuffle + decryption proofs to the transcript.
• Audit — the verifier re-checks the whole transcript (RunVerify): every CC Schnorr proof and the full shuffle chain, from public data alone.

Design decision: proofs on the bulletin board, ciphertexts on the wire

Fully serializing the Bayer-Groth argument tree (product / Hadamard / zero / SVP / multi-exponentiation sub-arguments, each with nested vectors) as JSON for every CC→CC hop would be enormous. Instead — matching the real system, where control components publish to a public bulletin board — the ciphertext handoffs between parties cross the signed transport (and are re-validated as G_q members on decode), while the zero-knowledge proofs are posted to the public transcript that the verifier independently re-checks. The vote data itself always crosses the authenticated channel; the proofs live on the board, exactly where an auditor reads them.

Trust-boundary hardening folded in here

Findings from the due-diligence pass that only bite once inputs are remote are addressed in the party layer:

• M4 (unvalidated deserialization): every crypto object is decoded through NewGqElement/NewZqElement at the wire boundary (pkg/party/wire.go), so a peer cannot inject a non-residue or out-of-range value.
• F6 (vcPK not persisted): the ballot carries vcPK, stored in the ballot box, so the exponentiation-proof statement is reconstructible by any party.
• F7 / F8 (padding / partial decryptions not persisted): the padded mix input and every per-stage partial decryption are published to the transcript, and the verifier checks the shuffle chain against exactly those.
• F12 (panic on malformed plaintext): tally uses DecodeVoteChecked, so a ballot that does not decode is spoiled, not fatal; handlers reject malformed messages cleanly rather than panicking.
• F5 / F16 (return codes were decorative): the return-code path is now genuinely cast-as-intended — see below.

Cast-as-intended return codes

The return code the voter checks is computed by the control components from the submitted ciphertext, so client-side vote substitution is detected:

1. The voter submits the ballot E1 plus a second ciphertext E2 = Enc(vote, returnCodesPK) and a plaintext-equality proof that E1 and E2 encrypt the same vote. Every CC verifies this proof; a ballot whose two ciphertexts disagree is rejected.
2. The card code for option i is the algebraic base prime_i^{Σ_j k_j}, where k_j is CC j's voter-specific return-code key. Setup builds it by combining per-CC shares; the same k_j is used at vote time (one derivation — the old F1 mismatch is gone by construction).
3. At vote time each CC exponentiates E2 by its k_j (product over CCs = Enc(vote^{Σk})), then contributes a partial-decryption factor. The server recovers vote^{Σk} — which equals prime_sel^{Σk} — and looks up the short code, returning it to the voter.
4. The voter checks the returned code against the card for the chosen option. A mismatch means the tallied vote differs from the intended one.

Soundness: because the equality proof binds E2 to E1 (the tallied ballot), the code is derived from the vote that is actually counted. Malware that alters the vote cannot produce a matching card code without the CC secrets. A unit test drives the substitution attack (E1=A, E2=B) and confirms the CCs reject it.

Remaining scope limits (honest)

• The return-code path is defined for a single selected option per voter (the demo case); multi-selection return codes are not implemented.
• Privacy: the return-code lookup lets the server learn which mapping entry matched, so in this PoC the return-code-computing parties can infer the selection. The real system separates these roles to keep the vote secret from any single party; that role separation is simplified here. Vote secrecy of the tally is unaffected (it rests on the mix-net, not the return-code channel).