2024 Sponge Example Analysis

This is the code for the paper The Impact of Uniform Inputs on Activation Sparsity and Energy-Latency Attacks in Computer Vision.

This script showcases how to use the strategies to generate sponge examples and measure their post-relu density as well as the amount of uniform surfaces.

Possible strategies are:

1. Random Image (Baseline)
2. Uniform Sampling Strategy
3. Natural Sampling (most densely activating image from ImageNet)
4. Sponge-GA Strategy
5. Sponge-LBFGS Strategy

Resulting images will be in the 'results' folder.

Please cite as follows:

@inproceedings{muequi2024,
      title={The Impact of Uniform Inputs on Activation Sparsity and Energy-Latency Attacks in Computer Vision}, 
      author={Andreas M\"{u}ller and Erwin Quiring},
      year={2024},
      booktitle={7th Deep Learning Security and Privacy Workshop (DLSP)},
}

Contact information:

• Andreas Müller
• Erwin Quiring

Setup

We used Python 3.9 for this project. Setup with conda:

conda create --name sponge_example_analysis python=3.9
conda activate sponge_example_analysis

Then install requirements:

pip install torch==1.12.1 torchvision==0.13.1 --index-url https://download.pytorch.org/whl/cu113 & pip install -r requirements.txt 

Then run

conda install jupyter 

This step is necessary for running Jupyter Notebook successfully on some systems.

Usage

Use either Jupyter Notebook and run example.ipynb or run

python example.py

They contain the same code.

Crediting

This project uses code from sponge_examples, authored by Ilia Shumailov. That code is located in the original_sponge_examples_code directory. The code is licensed under MIT License.

Please see the license of the sponge_examples repo in the original_sponge_examples_code/LICENSE-original file.

The sponge_examples repo contains code for the publication "Sponge Examples: Energy-Latency Attacks on Neural Networks". The paper can be found on here or alternatively on arxiv. Please do not forget to give the authors credit. To cite please use:

@inproceedings{shumailov2020sponge,
      title={Sponge Examples: Energy-Latency Attacks on Neural Networks}, 
      author={Ilia Shumailov and Yiren Zhao and Daniel Bates and Nicolas Papernot and Robert Mullins and Ross Anderson},
      year={2021},
      booktitle={6th IEEE European Symposium on Security and Privacy (EuroS\&P)},
}