Jingle-Back: Backdoor Attacks with Stylistic Transformations

In this project we implement a backdoor attack in speech recognition with stylistic transformations as our trigger. For our stylistic transformations we used Spotify's Pedalboard and we named our attack Jingle-Back.

Guide

First a virtual environment has to be created, and activated. Then the requirements should be installed:

$ python -m pip install -r requirements.txt

To download the speech commands dataset and calculate features for both the clean and the poisoned samples the script startup.sh should be run:

$ ./startup.sh

This script selects only 10 classes from the whole dataset to demonstrate the functionality faster in less data. For the data preprocessing prepare_data.py is used.

After preparing the features, the models can be trained with the following command:

$ python train.py mfccs data_10

About

This is the repo for our paper "Going in Style: Audio Backdoors Through Stylistic Transformations" which was presented in ICASSP 2023. To reference our paper use the following bibtex entry:

@inproceedings{koffas2023going,
  title={Going In Style: Audio Backdoors Through Stylistic Transformations},
  author={Koffas, Stefanos and Pajola, Luca and Picek, Stjepan and Conti, Mauro},
  booktitle={ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={1--5},
  year={2023},
  organization={IEEE}
}