What is a zero-knowledge proof?

A zero-knowledge proof lets someone prove that a statement is true without revealing anything else, for example that they own funds and may spend them, without showing which funds or who they are.

Also called ZK proof or ZKP.

Explain it like I’m five

A ring-shaped cave has a magic door across the middle that opens only with a secret word. Peggy says she knows the word, and Victor wants proof without learning it. Peggy walks in and takes the left or the right path while Victor waits outside, unable to see. Victor then shouts which side he wants her to come out of. If she knows the word, she opens the door when needed and comes out of the right side every time. If she does not, she succeeds only when she happened to pick that side, which is half the time.

After 20 rounds a bluffer gets through about once in a million times, exactly 1 in 1,048,576, and each further round halves that. Victor ends up convinced, and he has learned nothing about the word.

What problem it solves

Proving something usually means showing the evidence: a password, a passport, a transaction history. A zero-knowledge proof keeps the evidence with its owner and still convinces the other side. This matters on a public blockchain, where all nodes check the rules in the open and so usually see all the details. With zero-knowledge proofs the chain can check that money exists, belongs to the spender and has not been spent before, while the owner, the amount and the history stay private.

How it works, step by step

  • The prover and the verifier agree on a public statement, and the prover holds a secret that makes it true (the witness).
  • The prover picks a fresh random value and sends a sealed first message built from it (the commitment).
  • The verifier answers with a random challenge, chosen after it has seen that message.
  • The prover replies with a number that mixes the secret, the random value and the challenge. Only someone who knows the secret can answer whichever challenge arrives, and the fresh random value masks the secret inside the reply.
  • The verifier checks the reply against the statement and the first message. One round leaves a bluffer a real chance, so the exchange repeats, or uses a challenge from a huge range, until that chance is negligible.
  • To remove the live conversation, the challenge is replaced by a hash of the statement and the first message (the Fiat-Shamir heuristic). The prover then sends one proof that anyone can check later (a non-interactive proof).

A worked example

The public number is 8. Peggy claims she knows a secret x such that 5 raised to the power x, taken modulo 23, equals 8. Her secret is 6. She picks the fresh random number 3 and sends 5 to the power 3 modulo 23, which is 10. Victor replies with the challenge 2. Peggy answers 3 plus 2 times 6, which is 15. Victor checks that 5 to the power 15 equals 10 times 8 squared, both modulo 23, and both come to 19. The answer 15 says nothing about 6, because the fresh 3 masks it.

A prover who could answer two different challenges to the same first message could work out the secret from the two answers, so only a real holder of the secret keeps answering correctly. Anyone can also fake a finished transcript by choosing the challenge and the answer first and computing the first message backwards, which shows that the transcript carries no information about the secret.

What it guarantees

  • An honest prover who knows the secret always convinces an honest verifier (completeness).
  • A prover who does not know it succeeds only with a probability that the designers set as small as they like, such as 1 in 1,048,576 after 20 rounds of the cave (soundness).
  • The verifier learns nothing beyond the fact that the statement is true. Formally, a simulator that does not know the secret can produce fake transcripts that look like real ones (zero knowledge).
  • The likeness can be exact (perfect), nearly exact (statistical) or hold only against limited computing power (computational).
  • Against a prover with unlimited power the system is a proof, and against a prover with limited power it is an argument.

Common misunderstandings

  • Zero knowledge still reveals that the statement is true, and everything that follows from it.
  • A proof holds no encrypted copy of the secret, so nothing in it can be decrypted later.
  • Zero-knowledge and succinct are separate properties. The ethereum.org documentation notes that validity proofs used for scaling do not by themselves make transactions confidential.
  • The example above rests on discrete logarithms, which a large quantum computer could break. Proofs built from hashes, such as STARKs, avoid that.

Where it is used

Zcash brought zero-knowledge proofs to payments in 2016, building on the 2014 Zerocash design, whose transactions are under 1 kilobyte and verify in under 6 milliseconds. The ethereum.org documentation also lists private payments such as Tornado Cash, identity systems built on Semaphore, voting systems such as MACI and zk-rollups, which check thousands of transactions with one proof.

History

Shafi Goldwasser, Silvio Micali and Charles Rackoff introduced zero-knowledge proofs in 1985. Fiat and Shamir showed in 1986 how a hash can stand in for the verifier, Blum, Feldman and Micali showed in 1988 that a shared random string removes the interaction, and Jens Groth’s 2016 system shrank a proof to three curve points. STARKs followed in 2018.

In Redacted

Every spend from your private balance is backed by a zero-knowledge proof made in your browser. It shows the contract that you own notes in the Reserve worth enough for the action, never which notes they are, and the contract checks every proof on-chain, in public, every time.

How privacy works →

Related terms

Sources: The Knowledge Complexity of Interactive Proof Systems, Goldwasser, Micali and Rackoff, SIAM Journal on Computing, 1989 · Zero-knowledge proof, Wikipedia · Zero-knowledge proofs, ethereum.org · Zerocash: Decentralized Anonymous Payments from Bitcoin, Ben-Sasson et al., 2014 · Zero-knowledge rollups, ethereum.org

Updated November 1, 2026. All terms