Paper 2026/840

All You Need Is Addition

Dimitrios Schoinianakis, Nokia Bell Labs
Abstract

A practical acceleration framework for CKKS homomorphic encryption is proposed, in which multiplication-heavy sub-circuits are evaluated by adding encrypted log-magnitudes rather than multiplying ciphertexts. This logarithmic-number-system (LNS) representation consumes no multiplicative levels; a lightweight interactive refresh operation re-enters the linear CKKS domain whenever additive accumulation is required, avoiding bootstrapping entirely. Three execution strategies---an automatic planner, a client-side accumulation variant, and a server-side re-encryption variant---are mechanized in an OpenFHE-based runtime under 128-bit classical security and evaluated on a deep multiplication chain and an attention-like pipeline across three network environments. On the attention pipeline, the linear baseline requires a ring degree of 65,536 and about 2.9 GB of public-context material, whereas LNS operates at a ring degree of 8,192 with about 50 MB---a 58x context reduction yielding 22-36x end-to-end speedup and up to 46x payload reduction. On the multiplication chain the gap widens with depth, from about 10x speedup and 26x payload reduction at L=8 matrices, to 35x and 98x at L=20, because LNS context requirements are decoupled from multiplicative depth while the linear baseline must escalate its ring degree to maintain accuracy and security.

Metadata
Available format(s)
PDF
Category
Cryptographic protocols
Publication info
Preprint.
Keywords
CKKShomomorphic encryptionlogarithmic number systemsecure inference
Contact author(s)
dimitrios schoinianakis @ nokia-bell-labs com
History
2026-05-03: approved
2026-04-29: received
See all versions
Short URL
https://ia.cr/2026/840
License
Creative Commons Attribution-NonCommercial-NoDerivs
CC BY-NC-ND

BibTeX

@misc{cryptoeprint:2026/840,
      author = {Dimitrios Schoinianakis},
      title = {All You Need Is Addition},
      howpublished = {Cryptology {ePrint} Archive, Paper 2026/840},
      year = {2026},
      url = {https://eprint.iacr.org/2026/840}
}
Note: In order to protect the privacy of readers, eprint.iacr.org does not use cookies or embedded third party content.