pllmR21: NEXUS: Secure Transformer Inference Made Non-interactive
NEXUS combines approximate HE with lightweight garbled circuits for non-interactive two-party transformer forward inference. PLLM status: tracked.
Status: tracked
Year: 2025
Authors: Jiawen Zhang, Xinpeng Yang, Lipeng He, Kejia Chen, Wen-jie Lu, Yinghao Wang, Xiaoyang Hou, Jian Liu, Kui Ren, Xiaohu Yang
External source: NEXUS: Secure Transformer Inference Made Non-interactive
Source contribution
NEXUS combines approximate HE with lightweight garbled circuits for non-interactive two-party transformer forward inference.
PLLM perspective
Reproduce the exact artifact first, then measure its operator components against the complete PLLM Qwen decoder graph without inferring KV-cache or token-feedback support.
No PLLM NEXUS reproduction has been executed.
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: Two-party semi-honest security and one inference evaluator do not establish complete private autoregressive LLM serving.
PLLM modules and documentation
Target module identifiers: protocol.nexus, gate.nexus_nonlinear, kernel.nexus_he. These identifiers describe scoped work unless this page explicitly links an implementation source. Dependencies: None.
Backlog
This source is assigned to reimplementation priority 8, Full-system comparison targets. Its source-registry order is 21. Follow current gates and unfinished work in the research backlog.
R20: Oblivious Ciphertext Compression via Linear Codes
This work develops oblivious ciphertext compression using linear codes and syndrome decoding. PLLM status: tracked.
R22: BumbleBee: Secure Two-party Inference Framework for Large Transformers
BumbleBee implements secure two-party transformer inference with specialized nonlinear protocols and an Apache-2.0 artifact. PLLM status: tracked.