Wiki Home
Cryptographic commitment
Definition
A commitment scheme lets a party publish a short token, the commitment, that fixes a value without revealing it, and later open the token to prove which value was committed. The scheme is hiding (the commitment leaks nothing about the value) and binding (the committer cannot later open it to a different value). A common instantiation is a cryptographic hash of the value together with a random nonce.
Publishing a commitment on a public, time-stamped, append-only log, a bulletin board such as a blockchain, adds proof of anteriority: evidence that the committed value existed no later than the post. In model-ownership settings this lets an owner commit to a model, and to per-claim secrets, before any dispute arises, so that a later claimant cannot fabricate an earlier-looking claim. The commitment carries the proof of priority; a separate trusted party or protocol checks the opened values against a suspected model.
Papers that use this concept
- DAWN: Dynamic Adversarial Watermarking of Neural Networks — the owner registers commitments to the protected model and to each client’s watermark on a public bulletin board, so a trusted judge can later confirm ownership and its priority in time.
- zkGPT: An Efficient Non-interactive Zero-knowledge Proof Framework for LLM Inference — a polynomial commitment to the model weights fixes one model in advance, so the zero-knowledge proof certifies inference under the committed weights and the prover cannot swap models between queries.
- PAL*M: Property Attestation for Large Generative Models — notes the cryptographic-attestation alternative, where publicly verifiable commitments bind a model’s training and inference for auditing, in contrast to its own hardware-rooted measurements.