# BioBind: Binding Biometric Deduplication to Zero-Knowledge Authentication

Crypto.world

> Binds enrollment deduplication to zero-knowledge authentication so concurrent enrollments cannot mint duplicate credentials, with released code an implementer can actually test for the race.

**arXiv**: 2611.12636

**Published**: 2026-11-09

**Authors**: Taotao Wang, Xiaoxiao Wu, Shengli Zhang

## AI summary

## Overview

**Research area:** Applied cryptography and security for decentralized identity — specifically biometric deduplication (1:N), zero-knowledge authentication, and smart-contract credential issuance.

**Technical level:** Advanced. The paper assumes familiarity with homomorphic encryption (BGV), threshold committees, zero-knowledge proof circuits, commitment schemes, and game-based security definitions.

**Scope (1 sentence):** The paper specifies four interface conditions that bind an off-chain private biometric deduplication verdict to an on-chain gallery update and credential, then presents and evaluates an instance built on BioZKFHE plus a capture-bound zero-knowledge authentication proof.

## What This Paper Is About

Decentralized identity systems want "one person, one account" without revealing who a user is, so they split the work: biometric capture and large private 1:N deduplication run off chain, while a smart contract authorizes credential issuance and later use. The gap is that a *correct* deduplication verdict only covers the specific inputs checked and the particular gallery snapshot queried, so a malicious enrollee running concurrent sessions, or a misbehaving backend compute node, can turn correct verdicts into duplicate credentials. BioBind's goal is to make explicit and checkable which captured vector was queried, which entry was written into the gallery, which commitment was issued, and whether the unique verdict still held when the entry was actually added.

## Key Contributions

1. **Four enrollment composition conditions (I1–I4).** Input binding — the queried, inserted and committed templates are one captured vector, with an enrollment-station link from the pending record to the credential commitment; authenticated gallery transition — an accepted update appends exactly the declared template at the slot assigned from the authenticated count; verdict linearization — a session may finalize as unique only if no write occurred after the snapshot it queried; and verdict soundness — a unique verdict is correct for the queried template and snapshot under the backend's corruption model. Counterexamples show what breaks when each condition is dropped.
2. **A construction and conditional security argument.** For a BioZKFHE-based backend, requests are bound to one captured vector at request time, sessions whose gallery snapshot has gone stale are rejected, and the committee confirms each merge of a new ciphertext into the gallery. Capture-bound zero-knowledge authentication links credential use to the enrolled template and a fresh device-originated probe bound to the challenge. A multi-session composition argument sorts failures into cryptographic ones, trusted-device assumptions, state invariants and biometric errors.
3. **Two authentication privacy profiles.** A public profile exposes holder address, credential handle, enrollment commitment, device key and a per-device counter on chain; an unlinkable profile keeps these in the zero-knowledge witness and uses per-application nullifiers derived from a holder secret that never reaches the enrollment station.
4. **Implementation and evaluation on real biometric features.** Contracts, circuits and an enrollment bridge tested against the composition attacks (each paired with an honest run), an end-to-end pipeline on LFW face features, an equal-public-field baseline isolating the cost of verifiable matching, and separate component tests of real BGV encryption over three histories of 16,384 registrations.

## Main Findings

- **Two concrete attacks motivate the design.** In a *snapshot race*, one person opens two enrollment sessions against a snapshot containing no entry of theirs; both queries correctly return unique, and both sessions insert and issue, so the person enrolls twice. In a *dropped merge*, a compute node discards the new ciphertext while reporting an updated block commitment, leaving the person absent from the gallery. Both are reproduced on a plaintext model of BioZKFHE's gallery encoding.
- **The unlinkable authentication circuit is small enough for on-chain verification.** It uses 31,445 constraints and approximately 305,000 gas.
- **Keeping matching verifiable has a measurable cost.** Against an equal-public-field baseline where the capture device attests to the match instead of the circuit proving it, witness-plus-proving medians are 1.989 s versus 1.119 s.
- **Encrypted enrollment ran at scale in component tests.** Three histories of 16,384 registrations completed without decoding errors at the tested checkpoints; the end-to-end pipeline uses a simulated committee.
- **Biometric accuracy bounds person-level uniqueness.** On the Bosphorus database, cooperative subjects who change their appearance evade deduplication in 1.8% of re-enrollment attempts at the operating threshold, and the end-to-end run records one missed duplicate.
- **Authentication needs no online committee or matching service.** BioBind takes the route of zero-knowledge 1:1 distance checks against a committed template, keeping raw features on the user side so the comparison itself is publicly verifiable; the committee's duty extends from confidentiality to confirming gallery-root migration, with the same party set and corruption threshold.
- **Person-level guarantees are conditional on attempt budgets.** The uniqueness bound is Adv_unq + q_E·f_nm and the impersonation bound is Adv_auth + B·f_max when the authentication capture module enforces its budget; f_nm and f_max are taken under adversarial presentation, so measured FNMR and FAR are only non-adversarial reference values.
- **Unlinkability is conditional at the protocol level.** Under the stated zero-knowledge, commitment-hiding and pseudorandomness assumptions the observer advantage is negligible, but the claim excludes network identification, compromise of a holder's host or device, differing metadata between candidates, and anonymity from fresh addresses alone.
- **The probe is not hidden from a malicious host.** Because the host runs the prover, it receives the probe opening; hiding it would require an attested prover reached over an encrypted channel.
- **Scope limitation stated by the authors.** Full backend integration remains open; corrupted devices and malicious owners of credentials, devices or policies are separate threat profiles not covered.

## Methodology in Plain English

The authors start by treating enrollment as a distributed-systems problem rather than a matching problem: capture, private search over a large gallery, and on-chain issuance sit in different trust domains, and the enrollee controls its own host, its submitted data and its timing. They define what "composable under malicious enrollees" means as four interface conditions, then test each by removing it and constructing an adversary that wins — the counterexample table pairs each removed premise with the adversary's action and the violated winning condition.

They formalize the goals as games. An enrollment uniqueness game lets the adversary run many concurrent sessions and wins if a credential's commitment does not come from the deduplicated vector (W1), if the decrypted and decoded gallery diverges from the ideal gallery (W2), or if a credential is issued to someone already in the gallery immediately before its atomic insertion (W3) — evaluated at finalization, not at the queried snapshot, which is exactly what the snapshot race exploits. An authentication game covers credential origin (U1), capture origin (U2), a genuine whole match between the two committed vectors (U3), and live state at the linearization point (U4). Separate events capture person-level Sybil and impersonation, and a third game captures unlinkability for an on-chain observer.

The instance builds on BioZKFHE's BGV-based gallery encoding and verifiable similarity evaluation, which assumes an honest client; a certified, attended enrollment station plays that role and derives the query ciphertext, the write ciphertext and the enrollment commitment from a single captured, quantized vector, signing a declaration before the backend runs and a link signature after it. A certified authentication capture module commits to a fresh probe under the challenge, consumes one unit of an attempt budget and signs, so the untrusted host can only run the prover, not choose the probe. The holder's host then proves in zero knowledge that the two committed vectors are within the threshold squared Euclidean distance, and the gateway verifies the proof and checks current credential, policy and device state.

Evaluation combines real LFW face features in an end-to-end pipeline with a simulated committee, component tests that push real BGV encryption through 16,384-registration histories, and evasive re-enrollment experiments on the Bosphorus database to connect matcher accuracy to the person-level bounds.

## Why This Matters

For research, the paper reframes biometric proof-of-personhood as a composition problem across trust domains. Its four conditions are stated as an interface that any 1:N backend could be held to — the authors even sketch how they map onto World ID's secret-sharing protocol, where verdict soundness rests on a different trust model — and it separates cryptographic failures from trusted-device assumptions, state invariants and biometric error rates instead of lumping them together.

Real-world applications:

- **One-per-person benefit distribution** — aid, tokens or airdrops where duplicate enrollment directly inflates cost.
- **Sybil-resistant governance** — one-vote-per-person mechanisms in DAOs and other on-chain organizations that currently rely on token or address weighting.
- **Private biometric login to decentralized applications** — repeated authentication without an online matching service and without exposing templates on chain.
- **Humanitarian and credentialing deployments** — settings such as those studied by Wang et al., where unique registration is assumed before private token use, and where that assumption needs enforcement.

Industry relevance is direct for teams building decentralized identity credentials, smart-contract wallets with personhood requirements, and biometric enrollment pipelines that already lean on homomorphic encryption or MPC: the contribution is not a new matcher but a specification of what a contract may rely on, plus measured costs (31,445 constraints, roughly 305,000 gas, and a 1.989 s versus 1.119 s proving median) that make the trade-off between verifiable and device-attested matching concrete.

## Future Directions

- **Full backend integration** — the authors state this remains open; the end-to-end pipeline currently uses a simulated committee rather than a deployed one.
- **Porting the conditions to other backends** — the four conditions are mapped onto an MPC-style backend in the supplementary material, but whether real MPC-based systems such as World ID's protocol can satisfy I1–I4 in practice is left as a trust-model question.
- **Deployment parameters for attempt budgets** — the uniqueness bound depends on q_E, the number of enrollment captures summed over adversarial persons, which is treated as a deployment parameter that an attended station must enforce; how to set and audit it is unresolved.
- **Threat profiles outside the model** — corrupted capture devices and malicious owners of credentials, devices or policies are explicitly excluded, as is hiding the probe from a malicious host, which would need an attested prover reached over an encrypted channel.

## Target Audience

Researchers and protocol engineers working on decentralized identity, proof-of-personhood and Sybil resistance; practitioners building privacy-preserving biometric enrollment with homomorphic encryption, MPC or zero-knowledge proofs; and smart-contract security engineers who need to reason about what an on-chain contract can actually rely on when an off-chain verdict authorizes issuance. Readers without a background in applied cryptography will find the trust-model discussion accessible but the security definitions and construction demanding.

## Authors’ abstract

Biometric identity systems for decentralized applications need credentials that resist repeated enrollment and support private, publicly verifiable authentication. Capture and large-scale private deduplication run off chain while smart contracts authorize issuance and use; these stages span distinct trust domains, and a correct verdict covers only the checked inputs and gallery snapshot. Without binding and state checks, concurrent enrollments can issue duplicate credentials…

[The full abstract on arXiv](https://arxiv.org/abs/2611.12636)

[Read the original paper](https://arxiv.org/abs/2611.12636)

---

Canonical: https://crypto.world/paper/biobind-binding-biometric-deduplication-to-zero-knowledge-authentication-2611.12636
