pllmR18: Compact: Approximating Complex Activation Functions for Secure Computation
Compact builds input-density-aware piecewise polynomial approximations for complex activations. PLLM status: tracked.
Status: tracked
Year: 2024
Authors: Mazharul Islam, Sunpreet S. Arora, Rahul Chatterjee, Peter Rindal, Maliheh Shirvanian
External source: Compact: Approximating Complex Activation Functions for Secure Computation
Source contribution
Compact builds input-density-aware piecewise polynomial approximations for complex activations.
PLLM perspective
Implement its fitting objective and published approximation profile separately from protected evaluation. Freeze coefficient/interval hashes and add explicit tail policy.
Our Fourier/boundary-matched fits are distinct hypotheses, not Compact reproductions.
This paper remains tracked. Named modules below are implementation targets, not finished PLLM components; no reproduction, integration, assurance, or matched benchmark is claimed.
The source's boundary remains binding: Calibration is not a proof of a worst-case range. Model quality and exact agreement with the chosen approximation are distinct.
PLLM modules and documentation
Target module identifiers: numeric.piecewise_fit, compiler.approximation_profile. These identifiers describe scoped work unless this page explicitly links an implementation source. Dependencies: None.
Backlog
This source is assigned to reimplementation priority 2, Activation approximation. Its source-registry order is 18. Follow current gates and unfinished work in the research backlog.
R17: Curl: Private LLMs through Wavelet-Encoded Look-Up Tables
Curl uses wavelet-encoded lookup tables for private nonlinear operations and discusses probabilistic truncation security. PLLM status: tracked.
R19: Private and Verifiable Outsourcing of Open-Weight LLM Inference
The abstract describes private and verifiable open-weight inference with two malicious, non-colluding servers and model-owner-published information. PLLM status: tracked.