smoothingASR

Randomized smoothing adversarial defense for ASR models, with original enhancement and voting strategies to limit the drop in performance. This code goes along with the EMNLP 2021 article "Sequential Randomized Smoothing for Adversarially Robust Speech Recognition".

Setup

We use armory to run adversarial experiments in a controlled environment. Please refer to their README for setup.

To install all requirements we use a custom Docker image. You can set it up with command docker build -t smoothing/pytorch-asr --build-arg armory_version=0.13.4 docker/

The tag is important, as it is refered to in our config files.

If you do not wish to use docker or armory, please refer to the Dockerfile for all package and libraries installation commands.

Run

Using armory you can run any of our config file. For instance : armory run configs/pgd/10/g1_trained_rover.json --num-eval-batches 100

Or write your own config files for custom experiments.

Models

You will find all our pretrained models and some auxiliary files here. Dump them in the saved models folder you setup when configuring armory.

Generate examples

The export_samples field in armory configuration files lets you export audio adversarial examples. Here we share some files, generated with the attacks mentioned in our paper against our proposed defense.

Cite

If you use this code please cite our paper:

@inproceedings{olivier-raj-2021-sequential,
    title = "Sequential Randomized Smoothing for Adversarially Robust Speech Recognition",
    author = "Olivier, Raphael  and
      Raj, Bhiksha",
    booktitle = "Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2021",
    address = "Online and Punta Cana, Dominican Republic",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.emnlp-main.514",
    pages = "6372--6386",
}