New fully homomorphic encryption research: schemes, bootstrapping, hardware acceleration and encrypted machine learning — with benchmarks, not just ideas.
Five items this week, spanning TFHE bootstrapping noise analysis, a hardware-oriented NTT redesign for CKKS and BGV/BFV, and encrypted machine learning from GPT-2 decoding to reinforcement learning and R-based statistics. Most worth reading is the static bootstrap placement paper, which reports concrete H100 timings for server-side token generation. The NTT paper is the one to read for hardware, since it removes the need for a dedicated automorphism unit.
Compares single-value and multi-value programmable bootstrapping in FHEW/TFHE under one noise model and derives the exact noise amplification as a ratio of two variances. In the spherical model the canonical cofactor costs p(p+1)/6 noise against 2p for the pivot, and both are measured on a public implementation.
AR-HE runs the whole language-model generation loop on the server, selecting each token under encryption, with one rule that places every bootstrap and makes bootstrap cost a formula in context length. A GPT-2 small token costs 544 s on one H100 versus 4715 s unoptimized, the prompt step costs 4630 s, and the encrypted cache cuts a generated step from 11751 bootstraps to 1072.
Gives in-place butterfly algorithms for the negacyclic NTT and its inverse that output evaluations in an automorphism-compatible order, differing from Cooley-Tukey and Gentleman-Sande only in twiddle factors. Galois automorphisms on double-CRT ciphertexts then become a cyclic shift or adjacent-pair swap, removing the need for a dedicated automorphism unit in FHE accelerators for CKKS and BGV/BFV. No implementation or performance numbers are given in the text.
Streams that already watch this field. Follow one as it is — it costs nothing extra.
Everything in fully homomorphic encryption: new papers and benchmarks, libraries and releases, new projects and startups, funding, products, standards, talks and events, attacks and incidents — across research, industry and onchain FHE.
Everything around FHERMA, the FHE challenge platform by Fair Math: new challenges and prize pools, results and winners, solution write-ups, Polycircuit and OpenFHE-rs releases, partnerships and mentions.
Every new release of the homomorphic encryption libraries, compilers and SDKs — TFHE-rs, Concrete, fhEVM, OpenFHE, SEAL, HEIR, Lattigo, Poulpy, CryptoLab HEaaN and enVector, Desilo, IBM HElayers, Apple Swift HE, TenSEAL and GPU libraries — with what changed.