Over-the-Air Ensemble Inference with Model Privacy [aRxiv] [IEEE]

Please see an extended version of this work titled "Private Collaborative Edge Inference via Over-the-Air Computation", published in IEEE Transactions on Machine Learning in Communications and Networking (TMLCN): [arXiv] [IEEE] [GitHub]

This repository contains the source code for the "Over-the-air ensemble inference with model privacy" paper.

Citation

Please cite the paper if this code or paper has been useful to you:

@inproceedings{yilmaz2022over,
  title={Over-the-air ensemble inference with model privacy},
  author={Yilmaz, Selim F and Has{\i}rc{\i}o{\u{g}}lu, Burak and G{\"u}nd{\"u}z, Deniz},
  booktitle={2022 IEEE International Symposium on Information Theory (ISIT)},
  pages={1265--1270},
  year={2022}
}

Installation

• Install conda and torch manually (recommended)
• pip install -r requirements.txt

Running

• First train and cache the device models.
• Then you can generate figures, tables or run raw experiments.

Training CV models

• python train.py --data <data_name> --num_repeats 10 --num_devices 20 --num_epochs 50
• <data_name> can be cifar10, cifar100, mnist, fashionmnist

Training NLP models

• python nlp_train.py --data <data_name> --num_repeats 10 --num_devices 20
• <data_name> can be yelp_review_full, yelp_polarity, imdb, emotion

Running an Experiment

• See the bottom of ota_private_ensemble.py

Generating TeX Code for the Comparison table

• Run python figure_comparison_table.py

### Generate TeX Code for the Varying Conditions pgfplot

• Run python figure_conditions.py