ENERGY BACKDOOR ATTACK TO DEEP NEURAL NETWORKS

This code provides a PyTorch implementation of the paper titled ENERGY BACKDOOR ATTACK TO DEEP NEURAL NETWORKS.

The figure above provides an overview of the backdoored model. Neurons circled in orange refer to unnecessary neurons that fire when the trigger is present in the input.

Dependencies and Reproducibility

All dependencies can be found in file

We run our code in a singularity container that can be obtained using the following command

singularity build --fakeroot ./test_image.sif ./test_image.def

The content of the test_image.sif should be:

Bootstrap: docker
From: pytorch/pytorch:latest

%files
    dependencies.txt dependencies.txt

%post
    apt-get -y update && apt-get install -y python    
    pip install efficientnet-pytorch==0.7.1
    pip install dill timm tensorboard
    apt-get install -y < dependencies.txt

%runscript
    python -c 'print("Image successfully loaded!")'

But the provided files can also be run in any other environment with the required packages installed.

Acknowledgements

We use in our project:

• ASIC simulator developed in .
• L0 estimation's implementation and energy estimation functions provided in for our model's optimization.
• This project is funded by Région Bretagne (Brittany region), France, CREACH Labs and Direction Générale de l’Armement (DGA).