Threshold Encryption
An encryption scheme where the power to decrypt is split across several parties, and only a defined quorum acting together can recover the data. No single party can decrypt alone, and the loss of a few parties does not lock the data away.
Also known as: threshold decryption, threshold cryptography
Threshold encryption splits the ability to decrypt data across a group of parties so that no one of them can read it alone. A decryption key is divided into shares, and only when a defined threshold of shareholders cooperate, say three of five, can the plaintext be recovered. This removes the single point of failure that plagues ordinary key management: there is no one server, person, or device whose compromise hands an attacker the data, and no single custodian whose disappearance loses it forever.
The mechanism is closely related to secret sharing and multi-party computation. Each key server holds a share and, on a valid request, contributes a partial decryption; the partials combine into the full result only if enough are supplied. Crucially, the shares are never assembled into a single master key in one place, so the trust is genuinely distributed rather than merely delegated. Access can be made conditional: a well-designed system checks a policy before releasing its share, so decryption happens only when agreed rules are met.
For DeAI this is the primitive that makes decentralised storage private without reintroducing a trusted gatekeeper. Walrus stores data publicly by default and relies on a separate layer called Seal, which uses identity-based and threshold encryption: independent key servers each hold a share and release just-in-time decryption keys only when an on-chain access policy is satisfied. The developer chooses which servers form the committee, so the trust assumption is explicit rather than hidden. It lets sensitive model inputs, datasets, or agent memory sit on open infrastructure while keeping the power to read them spread across parties that must agree.