Here you can find the models, the datasets information and the code used in our experiments (Copycat and Copycat Expansion). Feel free to contact me for any questions or suggestions (jacsonrcsilva at gmail).
Note that we used the Caffe Framework (1, 2). Therefore, you will find the “prototxt” files to replicate our experiments.
But if you don’t want to use Caffe, it is not a problem. In order to make it easier for you, we are also providing the following codes implemented in PyTorch:
Also, if you want to see an interactive comparison between Oracle and Copycat models, visit: Copycat CNN Explainer
It is implemented in TensorflowJS, using the CNN Explainer system.
If something here was useful to you, please kindly cite our article (s) below.
😊
This paper is available on arXiv
@inproceedings{Correia-Silva-IJCNN2018,
author={Jacson Rodrigues {Correia-Silva} and Rodrigo F. {Berriel} and Claudine {Badue} and Alberto F. {de Souza} and Thiago {Oliveira-Santos}},
booktitle={2018 International Joint Conference on Neural Networks (IJCNN)},
title={Copycat {CNN}: Stealing Knowledge by Persuading Confession with Random Non-Labeled Data},
year={2018},
pages={1-8},
doi={10.1109/IJCNN.2018.8489592},
ISSN={2161-4407},
month={July}
}
This paper is available on arXiv
@article{Correia-Silva-PATREC2021,
author={Jacson Rodrigues {Correia-Silva} and Rodrigo F. {Berriel} and Claudine {Badue} and Alberto F. {De Souza} and Thiago {Oliveira-Santos}},
title={Copycat {CNN}: Are random non-Labeled data enough to steal knowledge from black-box models?},
journal={Pattern Recognition},
volume={113},
pages={107830},
year={2021},
issn={0031-3203}
}
Example Code for Copycat in PyTorch
The PyTorch Weights for Oracle and Copycat models can be downloaded here
Copycat CNN: Convolutional Neural Network Extraction Attack with Unlabeled Natural Images
(more details)
(2nd download option)
The PyTorch Weights for Oracle and Copycat models can be downloaded here
@phdthesis{correia-silva-phd-2023,
author = {Correia-Silva, Jacson Rodrigues},
title = {Copycat CNN: Convolutional Neural Network Extraction Attack with Unlabeled Natural Images},
year = {2023},
school = {Universidade Federal do Esp\'{i}rito Santo},
address = {Esp\'{i}rito Santo, Brazil},
}