Contents
StandardNIST

FIPS 203 (ML-KEM)

Module-Lattice-Based Key-Encapsulation Mechanism Standard

FIPS 203 defines ML-KEM, the NIST post-quantum way for two parties to establish a shared secret over a public channel.

What it is

FIPS 203 is the NIST standard for ML-KEM, a module-lattice key-encapsulation mechanism derived from CRYSTALS-Kyber. Two parties use it to agree a shared secret on a public channel. The sender combines the recipient’s public encapsulation key with fresh randomness and obtains both a ciphertext and a shared secret. The recipient applies the private decapsulation key to that ciphertext and recovers the same secret. Ordinary key derivation then turns the secret into symmetric keys. ML-KEM does not encrypt a chosen message, and it does not create signatures.

The algorithm is aimed at adversaries who store traffic and later attack it with a quantum computer. Vehicles keep diagnostic logs, backend sessions, and fleet archives long enough for that harvest-now, decrypt-later risk to matter. A bad ciphertext must not hand an attacker a distinct yes-or-no result to probe. Decapsulation still returns a secret-shaped value, and that value will not match the sender.

Who it applies to

U.S. federal systems use FIPS 203 when they need an approved post-quantum key-encapsulation mechanism. Vehicle type-approval rules do not name it. ISO/SAE 21434 leaves the algorithm to the risk treatment. OEMs and suppliers still specify ML-KEM when a NIST-approved KEM is the acceptance bar for cloud links, workshop backends, remote diagnostics, and update channels that must stay confidential for years. Module vendors implement it for NIST cryptographic-module validation. Teams who only configure SecOC still need the ciphertext size so a handshake is not placed on a bus that cannot carry it.

What it covers

The standard defines three parameter sets and three operations: key generation, encapsulation, and decapsulation. It fixes byte encodings, the randomness each operation requires, and the internal checks. Certificates, protocol framing, and bus schedules are out of scope.

Parameter set Security category Encapsulation key Ciphertext Shared secret
ML-KEM-512 1 800 bytes 768 bytes 32 bytes
ML-KEM-768 3 1184 bytes 1088 bytes 32 bytes
ML-KEM-1024 5 1568 bytes 1568 bytes 32 bytes

The shared-secret length stays 32 bytes. Public keys and ciphertexts grow with the category. Private decapsulation keys are larger still and are long-term secrets.

How it fits with other documents

FIPS 204 (ML-DSA) and FIPS 205 (SLH-DSA) are the NIST post-quantum signature standards that accompany FIPS 203. A protected channel needs a way to agree keys and a way to authenticate identities; a KEM and a signature algorithm do different jobs. TLS and other profiles define how the ciphertext is carried and how the 32-byte secret becomes traffic keys. Combining ML-KEM with a classical elliptic-curve exchange is common, so the session survives if one of the two assumptions fails. That hybrid pattern is a profile decision; FIPS 203 does not require it. ISO/SAE 21434 is the process that records whether quantum-capable attackers are in scope. An AUTOSAR crypto stack can expose encapsulation only after a library or HSM actually implements these operations.

In practice

Fix the parameter set in the cybersecurity concept. If the profile does not name one, record an explicit choice: ML-KEM-768 is a frequent commercial default, and some high-assurance profiles require ML-KEM-1024. Dropping to ML-KEM-512 needs a written security rationale. Budget for the ciphertext sizes in the table: routine on Ethernet and on TLS to the cloud, a poor match for cyclic CAN or CAN FD. Draw encapsulation randomness from a cryptographic RNG, and qualify the supplier with the standard’s test vectors. Use a maintained library or HSM firmware rather than a first implementation of the lattice math on an application core, and keep decapsulation constant-time. Store private keys in a secure element when the ECU has one. Keep the KEM replaceable so a later parameter set or a successor algorithm does not force a new network architecture.

What it means for your discipline

  • CybersecurityCore

    You own whether ML-KEM is the key-agreement control, which parameter set is required, and whether the handshake is also combined with elliptic-curve Diffie–Hellman. Daily work is writing that choice into the cybersecurity concept and checking long-lived fleet and backend traffic against harvest-now, decrypt-later.

  • HPC & application softwareCore

    You integrate key generation, encapsulation, and decapsulation into the security library, TLS stack, or HSM interface, including randomness and constant-time decapsulation. You then pass the 32-byte shared secret into the key-derivation path the rest of the stack already uses.

  • E/E architecturePartial

    You place ML-KEM on links that can carry ciphertexts of several hundred to about 1,500 bytes, and you keep a crypto-agility path so the algorithm can change without a new topology. Lattice arithmetic stays with the crypto supplier.

  • Base softwarePartial

    You size crypto-job buffers and key storage for the encapsulation key, decapsulation key, and ciphertext of the chosen set. The service layer can offer the operation only when the driver or HSM firmware implements it.

  • Hardware & siliconPartial

    You choose host software, an accelerator, or an HSM for the three operations and budget RAM and latency for that parameter set. Private decapsulation keys need the same protected store as other long-term secrets.

  • Networks & communicationPartial

    You plan session setup on Ethernet, DoIP, and cloud TLS around public keys and ciphertexts far larger than an elliptic-curve share. ML-KEM is a handshake cost, not a payload for every cyclic CAN frame.

  • Diagnostics & lifecyclePartial

    You meet ML-KEM when a tester or remote-diagnostics path opens a TLS session instead of relying only on SecOC. UDS services stay the same; the gateway and tester must accept the larger handshake.

  • Software update & OTAPartial

    You use ML-KEM when the update channel needs it to protect session keys in transit. Package authenticity still comes from a signature algorithm, which this document does not provide.

No direct stake: Functional safety