Post-quantum migration is a software engineering programme, not a single algorithm swap. This guide explains where ML-KEM, ML-DSA and SLH-DSA fit, compares their practical trade-offs, and gives developers a reusable checklist for inventory, integration, testing and staged rollout.
Overview
Post-quantum cryptography (PQC) uses cryptographic constructions intended to resist attacks from both conventional and sufficiently capable quantum computers. The first implementation question is not “which algorithm is strongest?” It is “what cryptographic operation does this part of the system require?”
- ML-KEM is a key-establishment mechanism. It is designed to help two parties establish a shared secret over an untrusted network. It replaces or complements key-agreement mechanisms; it does not create digital signatures.
- ML-DSA is a general-purpose digital signature scheme. It can authenticate software, certificates, documents, API messages and other signed data.
- SLH-DSA is also a digital signature scheme, based on a different design approach. It is often considered where algorithm diversity, conservative assumptions or long-term signing requirements matter more than compact signatures.
These names refer to the standardised forms of the algorithms. Avoid treating an early research implementation, a vendor-specific name or a pre-standard variant as automatically interchangeable with the final standard. Confirm the exact parameter set, encoding, protocol support and library version before committing to an interface.
ML-KEM, ML-DSA and SLH-DSA compared
| Algorithm | Primary purpose | Size and performance considerations | Typical placement |
|---|---|---|---|
| ML-KEM | Establishing shared encryption keys | Public keys and ciphertexts are substantially larger than many traditional elliptic-curve equivalents. The three standard parameter sets have public-key sizes of 1,184, 1,952 and 2,592 bytes, with ciphertexts of 768, 1,088 and 1,568 bytes. | TLS or other secure transport, service-to-service key establishment, encrypted storage key wrapping |
| ML-DSA | General-purpose signatures | Signatures and public keys are larger than common classical signatures. Depending on the parameter set, public keys are 1,312, 1,952 or 2,592 bytes, and signatures are approximately 2,420, 3,309 or 4,627 bytes. | Certificates, software signing, API authentication, document and transaction signing |
| SLH-DSA | Hash-based signatures | Public keys are compact, but signatures can be much larger than ML-DSA signatures. Size and signing speed vary considerably by parameter set and security or performance profile. | Selected long-lived signing, archival verification and applications that value cryptographic design diversity |
Sizes above are useful planning figures, not a substitute for measuring the exact library and parameter set in your protocol. Larger keys and signatures affect certificates, handshake messages, database columns, queues, headers, hardware modules and bandwidth. A technically correct algorithm can still create operational problems if existing limits are too small.
Checklist by scenario
For encrypted network traffic
- Map where ephemeral or long-term key agreement occurs: browsers, APIs, VPNs, message brokers, service meshes and private links.
- Identify the protocol and termination point. A library may support ML-KEM while the reverse proxy, load balancer, certificate service or managed platform does not.
- Test a hybrid design where classical and post-quantum mechanisms are used together, if supported by the protocol and approved architecture. Define how the combined result is derived and how negotiation failures are handled.
- Measure handshake size, connection setup time, CPU use, memory, retransmissions and behaviour on constrained clients.
- Decide whether a failed post-quantum negotiation should fail closed, fall back to a classical mode temporarily, or route to a separately controlled endpoint. Document the decision rather than inheriting a library default.
For software, firmware and document signing
- Use ML-DSA as the starting candidate for general-purpose signing, then assess whether SLH-DSA is appropriate for a specific long-lived or diversity-sensitive use case.
- Check signature size limits in update manifests, package indexes, certificate chains, secure-boot storage and offline verification tools.
- Test key rotation, revocation, timestamping, rollback protection and recovery after a failed update.
- Preserve the algorithm identifier and parameter set in signed artefacts so verification is unambiguous.
- Keep a migration path for existing classical signatures. During a transition, verify both old and new formats only where the trust policy explicitly permits it.
For APIs and internal services
- Separate authentication from encryption. ML-DSA addresses signatures; ML-KEM addresses key establishment.
- Define whether signing happens at the client, gateway, service or message level. This determines key ownership and rotation responsibilities.
- Test headers, tokens, queues and database fields against larger keys and signatures.
- Record algorithm, parameter set, key identifier and verification result in audit data without logging private keys or sensitive payloads.
What to double-check
Standards and names: Confirm that documentation uses ML-KEM, ML-DSA and SLH-DSA as applicable, rather than relying on informal or pre-standard labels. Track the relevant NIST standards and your organisation’s approved cryptographic policy.
Library maturity: Check whether the implementation is production-ready, maintained, reviewed, constant-time where required, and backed by a clear vulnerability-response process. “Supported” may mean anything from an experimental API to a stable, interoperable implementation.
Protocol interoperability: Test both ends with the same parameter set and encoding. Include certificate chains, proxies, hardware security modules, managed cloud services and older clients in the test matrix.
Key management: Verify generation, storage, backup, rotation, destruction, access control and disaster recovery. A new algorithm does not compensate for poor private-key governance.
Performance: Measure real workflows rather than isolated primitive benchmarks. Record latency percentiles, throughput, memory, message size, CPU, failure rates and cost-relevant resource use.
Classical fallback: A fallback can preserve availability, but it can also quietly defeat the migration objective. Define who can enable it, for how long, with what monitoring and under which incident conditions.
Common mistakes
- Choosing one algorithm for every job: ML-KEM is not a signature scheme, and a signature scheme is not a key-establishment mechanism.
- Ignoring message expansion: Larger certificates and signatures can exceed fixed protocol, database or device limits.
- Testing only a successful path: Include malformed keys, unsupported parameter sets, truncated messages, timeouts, rotation and downgrade attempts.
- Assuming hybrid means automatically secure: Validate the protocol’s combination method and confirm that an attacker cannot force an unintended classical-only path.
- Replacing cryptographic calls without an inventory: Search source code, build dependencies, operating systems, appliances, vendor contracts, certificates, backups and archived data.
- Copying sample code into production: Examples may omit key protection, randomness requirements, error handling and lifecycle controls.
- Promising a permanent answer: Standards, implementations and protocol support can change. Record assumptions and make them reviewable.
When to revisit
Review the migration plan at least every six months and sooner when a relevant standard, library, protocol, hardware platform or supplier changes. Revisit it before seasonal planning cycles, major application releases, certificate renewals, fleet refreshes, cloud migrations and long-term data-retention decisions.
At each review, repeat this short action list:
- Update the cryptographic dependency inventory and mark newly discovered unknowns.
- Check official standards, implementation advisories and release notes for every selected library and protocol.
- Re-run interoperability and performance tests using production-like traffic and representative devices.
- Review fallback usage, failed negotiations, algorithm distribution and key-rotation evidence.
- Record whether each workload is ready, blocked, migrating or accepted temporarily as a risk.
The practical goal is not to declare a one-time “PQC complete” milestone. It is to make cryptographic dependencies visible, choose algorithms by function, test the full delivery path and retain enough flexibility to respond safely as the ecosystem develops.