4 articles tagged with private-inference.
ACM CCS 2025: diagonal representation + compact packing drive the rotation count down to 2%–5% of BOLT in the two most expensive CC-MM cases, and BERT-base with 128 tokens runs in 10 minutes on a single GPU; but 56% of a layer's time goes to bootstrapping.
Deep-read notes on arXiv 2412.08145: threat models, the division of labour between MPC and HE, the dispute over the 12.7% bottleneck, the routes for linear and non-linear layers, the two independent axes of evaluation, and errata E1–E14 in the original text.
Notes on arXiv 2609.12378: CKKS packing determines the cost of LLM homomorphic inference; the body of the paper's Chapter 4 is missing, the core evidence sits in the authors' open-source repository, and the 4.51× is not reproducible.
IEEE S&P 2024: a hybrid of MPC and HE that cuts Transformer inference communication to 1/10.91 and corrects the error in Iron's LayerNorm simplification.