Iron: Private Inference on Transformers

Authors: Meng Hao, Hongwei Li, Hanxiao Chen, Pengzhi Xing, Guowen Xu, Tianwei Zhang

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments conducted on several real-world datasets and models demonstrate that Iron achieves 3 14 less communication and 3 11 less runtime compared to the prior art.
Researcher Affiliation Academia Meng Hao1 Hongwei Li1 Hanxiao Chen1 Pengzhi Xing1 Guowen Xu2 Tianwei Zhang2 1University of Electronic Science and Technology of China 2Nanyang Technological University
Pseudocode Yes Algorithm 1 Secure Matrix Multiplication Protocol; Algorithm 2 Secure Softmax Protocol; Algorithm 3 Secure GELU Protocol; Algorithm 4 Secure Layer Norm Protocol.
Open Source Code No The paper states 'Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [No]' in its ethics checklist. While it mentions building on existing libraries, it does not provide the code for Iron itself.
Open Datasets Yes We train the models for four NLP tasks over the datasets of the Stanford Sentiment Treebank (SST-2), the Microsoft Research Paraphrase Corpus (MRPC), the Multi-Genre Natural Language Inference Corpus (MNLI) and the Stanford Question Answering Dataset (QNLI) from GLUE benchmarks [18].
Dataset Splits Yes We train the models for four NLP tasks over the datasets of the Stanford Sentiment Treebank (SST-2), the Microsoft Research Paraphrase Corpus (MRPC), the Multi-Genre Natural Language Inference Corpus (MNLI) and the Stanford Question Answering Dataset (QNLI) from GLUE benchmarks [18]. The GLUE benchmarks are standard datasets with predefined train, validation, and test splits.
Hardware Specification Yes All the following experiments are performed on AWS c5.9xlarge instances with Intel Xeon 8000 series CPUs at 3.6GHz.
Software Dependencies No Iron is built on top of the SEAL library [32] and the EMP toolkit [33] in C++. We also use the Ez PC framework [34]. No specific version numbers are provided for SEAL, EMP toolkit, or Ez PC.
Experiment Setup Yes These models are parameterized by three hyper-parameters: the number of blocks, the dimension of representations and the number of input tokens (refer to Appendix A.4.1 for the hyper-parameters of these models).