Source code of FlowMur

Jiahe Lan, Jie Wang, Baochen Yan, Zheng Yan and Elisa Bertino, "FlowMur: A Stealthy and Practical Audio Backdoor Attack with Limited Knowledge," IEEE S&P, 2024.

The workflow of FlowMur

How to start

This example is for the following setting:

dataset --> Google Speech Command Dataset V2

target model --> SmallCNN；surrogate model --> LargeCNN

#class of $D$ --> 10; #class of $D_{aux}$ --> 25; #class of $D_{sur}$ --> 26

Step 1: Data Preprocessing

Extract audio features for $D$ and $D_{sur}$ respectively.

python data_preprocessing.py

Step 2: Obtain the Surrogate Model

Train the surrogate model on $D_{sur}$

python Benign_Model.py

Step 3: Generate the Trigger

Optimize the trigger on the surrogate model

python generate_trigger.py

Step 4: Data Poisoning and Backdoor Injection

Poison $D$ and train SmallCNN on poisonous $D$

python Attack.py

How to cite

If you find this work useful, please consider citing it as follows:

@inproceedings{lan2024flowmur,
  title={FlowMur: A Stealthy and Practical Audio Backdoor Attack with Limited Knowledge},
  author={Lan, Jiahe and Wang, Jie and Yan, Baochen and Yan, Zheng and Bertino, Elisa},
  booktitle={2024 IEEE Symposium on Security and Privacy (SP)},
  pages={148--148},
  year={2024},
  organization={IEEE Computer Society}
}

Comments

If you have any questions about the code, please feel free to ask here or contact me via email at [email protected].