Skip to contents

Changelog

Source: NEWS.md

Version 1.0

CRAN release: 2026-10-03

Encryption backends

  • New OpenFHE backend, via the openfhe.R package. CKKS carries real-valued data natively, so the fixed-point denominator the Paillier path needs is gone; BFV and BGV carry exact integers. The scheme is read back from the crypto context, so one master drives any of them.
  • The Paillier implementation remains and is frozen: it is still exported and tested, but new work should use the OpenFHE backends.

Multi-party actors

  • Site is now abstract, with two concrete kinds. LocalSite holds its data in the current R session. RemoteSite is for a party whose contribution is produced elsewhere; no transport ships with the package, so subclass it and register a contribute() method.
  • contribute(site, theta) is the single call a master makes on a site. It returns the site’s contribution already encrypted, so no individual cleartext value reaches the aggregator.
  • Two failure modes are deliberately distinct: returning NA means theta is non-evaluable at a site that answered, while site_unavailable() signals that the site could not be reached and aborts the round rather than silently changing the set of sites being summed over.
  • make_worker(name, data, contribution_fn) replaces make_site(name, data, local_fn).
  • Masters: make_ckks_master() when one party may hold the secret key, make_threshold_master() when none may, and the frozen make_master() for Paillier.
  • master_aggregate(master, theta) runs the star (master/worker) topology that distcomp- and DataSHIELD-style analyses use. round_robin_chain() / run_round_robin() remain for the frozen Paillier chain idiom.

Threshold key generation

  • make_threshold_master(name, cc, sites) runs the key-generation chain through the sites: each generates its own secret share, keeps it, and passes on only a public key. The master holds the crypto context and the joint public key, and no secret material at all.
  • Sites must therefore be constructed before the master, since the joint public key is a function of all of them.
  • decrypt() recovers a value by asking every site for a partial decryption and fusing the results. No party, the master included, can decrypt alone.
  • New generics keygen_round() and partial_decrypt(), dispatching on Site. A RemoteSite subclass must implement both; the defaults refuse rather than generate a remote party’s share locally.
  • The help page states what the construction does not defend against: participants are assumed to follow the protocol, and a deviating site can return a dishonest contribution or a malformed partial that corrupts the result silently.

Setup, and what a remote site must implement

  • set_public_params(site, params) is the setup seam: the one moment a coordinating party hands a site anything, apart from a round itself. The Site method stores the bundle; the RemoteSite method refuses, because storing it would configure the local proxy and leave the far endpoint untold. Wiring a remote subclass that has not implemented provisioning now fails immediately instead of producing a site that looks configured and is not.
  • site_params(site) reads back what a site holds, and errors if it was never configured. Use it instead of reaching into site@state$params.
  • Public parameters are now typed S7 objects — PublicParams with OpenFHEParams and the frozen PaillierParams — rather than a list with a scheme string. “The bundle carries no secret material” is now a property of the class rather than a promise about a list, and the objects print to show it.
  • master_aggregate() checks each reply before adding it to a total: a site that answered in cleartext, or with a value produced under some other key, is refused. Previously a cleartext reply was folded in by ordinary scalar addition and the round returned the right answer, having been handed the one quantity the protocol exists to hide.
  • decrypt() and site-side partial_decrypt() likewise verify that a value belongs to this protocol’s key, using OpenFHE’s key tag. A site therefore refuses to apply its own share to a ciphertext from a protocol it did not join.
  • make_threshold_master() rejects a site listed twice, two sites with the same name, and a site already serving another protocol; a ceremony that fails part-way rolls back, leaving the sites it had visited clean enough to retry.
  • ?RemoteSite now separates three cases that were previously run together: a LocalSite demonstration models the protocol’s roles in one R session and is not a trust boundary; a single-decrypter deployment relies on honest execution rather than cryptography; only a remote threshold deployment gives a real party boundary, and only if your transport, authentication, and key storage provide one.

Encryption surface

  • encrypt() and decrypt() are openfhe.R’s generics, imported, extended, and re-exported. homomorpheR no longer defines generics of its own under these names; its protocol-actor methods and the frozen Paillier methods register on the upstream generics, so there is one method table per verb and the class of the first argument selects the layer. Argument names follow openfhe.R, which follows the OpenFHE C++ signatures; every call in this package is positional. See ?actor-encryption.
  • encrypt(site, value) is the one encryption entry point a party needs. A site holds its public parameters, so it does not fetch them and hand them back: the site is the whole of the argument. encrypt(params, value) serves anyone holding a bundle without being a site — a querier, say — and site_params(site) still hands one out.
  • A site is autonomous once configured. It is given its public parameters once, and from then on computes and encrypts without consulting anyone. There is deliberately no exported function that reaches from a site back to a master; the master-side public_params() is not exported either, since fetching it at encryption time would mean asking for something already held.
  • Consequently there is no encryption entry point taking a master: no master_encrypt(), and no encrypt() method on Master. Naming or taking one party would advertise a privilege that does not exist, and would invite site-side code to hold a master it has no use for.
  • decrypt(master, ciphertext, len) does take a master, and that asymmetry is the point: decryption is privileged, requiring secret material or the standing to convene every site, while encryption is not. It is vector-aware via len.
  • Under BFV and BGV a value the scheme cannot carry is now refused rather than coerced: a non-integer, a non-finite value, one outside R’s integer range, or one at or beyond half the plaintext modulus. as.integer() previously turned a contribution of 0.9 into 0 and reported a total of zero without a warning. The one thing no party can check is the total, which still wraps if it exceeds the modulus; ?ThresholdMaster says so.
  • make_ckks_master() requires a CKKS context. Exact-integer work goes through make_threshold_master(), which is scheme-agnostic by design.
  • Crypto contexts, key pairs, public keys, state environments, and contribution_fn are now typed S7 properties rather than class_any, and a party’s name must be a single non-empty string.

Data

  • DLBCL (235 patients: survival, subgroup, gene-expression signatures) and DLBCL_gex (235 x 6416 Lymphochip probes).
  • Precomputed results of the encrypted chunks, which a vignette build shows but does not run: cox_results, cox_threshold_results, cox_threshold_dp_results, cvxr_consensus, cvxr_admm_dp_results, and similarity_results. Each is produced by a data-raw/ script from its vignette’s own chunks; HOMOMORPHER_RECOMPUTE=true runs the chunks for real.

Vignettes

  • Thirteen vignettes covering queries and aggregation, model fitting across sites, prediction and retrieval, and Gaussian-noise variants; the Paillier-era vignettes have moved to paillier-archive/ and are no longer built.

Infrastructure

  • Migrated from R6 to S7 throughout. The R6 dependency is removed.
  • New PaillierCiphertext class wraps encrypted values together with the public key they were encrypted under. R’s arithmetic operators (+, -, * against a cleartext scalar) now dispatch directly on encrypted values via an S3 Ops group handler, so computations on encrypted data read like ordinary R arithmetic.
  • Paillier API renames:
    • PaillierKeyPair$new(bits) → paillier_keypair(modulus_bits)
    • pubkey$encrypt(m) → encrypt(pubkey, m) (S7 generic)
    • privkey$decrypt(ct) → decrypt(privkey, ct) (S7 generic)
    • pubkey$add(a, b) / pubkey$sub(a, b) → a + b / a - b
    • pubkey$mult(ct, k) → ct * k (cleartext scalar)
    • keys$getPrivateKey() → get_private_key(keys)
    • privkey$getLambda() → get_lambda(privkey)
  • Field access: keys$pubkey → keys@pubkey, etc.
  • Frozen Paillier-era code is separated into its own source file from the supported actor surface.
  • Minimum R version bumped to 4.3.0 (required for S7 @ and chooseOpsMethod). digest dropped from Suggests. Messaging goes through cli.

Version 0.3

CRAN release: 2025-04-09

  • Added MPC vignettes