PriFT

PriFT: Private Fine-Tuning using off-the-shelf MPC and HE libraries

Qiuxuan Ma, Eleftheria Makri, Nusa Zisaric

Venue
IACR ePrint 2026/1381
Date
2026-07-06
Affiliation
LIACS, Leiden University
Numbers from
primary
Uses ZK
no
Open source
yes
Paper
https://eprint.iacr.org/2026/1381

Reported benchmarks

modespeedup vs fully privateaccuracy retention
MPC semi-private3close to plaintext (no metric extracted)

Notes

Uses a transformer as a frozen feature extractor, then trains a small NN on the privacy-protected features. Two modes: fully-private (encrypted end-to-end) and semi-private (labels decrypted during training for speed). Findings: MPC beats HE, especially semi-private; MPC semi-private is ~3x faster than fully-private; accuracy close to plaintext. Its value here is as a clean MPC-vs-HE comparison on one real task. PDF is behind Cloudflare -- reference list NOT extracted; do that before trusting the adjacent-literature list below.

Our reading

Citation neighbourhood

Builds on
  • none recorded
Cited by, in this corpus
  • none recorded

Edges are a proxy: paper A's text mentions B anywhere (body or bibliography). See the full graph.

Discussed in

Other recorded fields
uses_zk_note: No proofs of any kind. PriFT provides PRIVACY (the data stays encrypted),
  not VERIFIABILITY (nobody proves the computation was correct). Orthogonal axis to
  this repo's other entries -- included to mark the boundary, not to compare.
primitive: MPC (Crypten) and HE (TenSEAL), benchmarked head-to-head