DeepTheft

This is the training source code for the offline phase of DeepTheft, which comprises two steps.

You can access the complete RAPL-based dataset at https://zenodo.org/records/14334986?preview=1&token=eyJhbGciOiJIUzUxMiJ9.eyJpZCI6IjFkNDNhNGM4LTI5NmYtNDVmYy04Njg0LTEyYmNlY2MzM2MzZiIsImRhdGEiOnt9LCJyYW5kb20iOiIyZjZmZDNhMWQxNWU2OTM5YTMyNWFiMzllY2NjZmY4OSJ9.m5FXwzGPETjAMh8_YUU4r7OkoFvb-Iaaj_ijDj-pMyYG5kjMwVNLdIxO94_MkUars4C9xeSOng_FMux2gGqLKA. The dataset includes "data.h5" and "hp.h5," representing the energy trace and hyperparameters, respectively. To execute the code successfully, please ensure that you place these files in the "datasets" folder within the root directory.

This is for releasing the source code of our work "DeepTheft: Stealing DNN Model Architectures through Power Side Channel". If you find it is useful and used for publication. Please kindly cite our work as:

@inproceedings{gao2024deeptheft,
  title={DeepTheft: Stealing DNN Model Architectures through Power Side Channel},
  author={Gao, Yansong and Qiu, Huming and Zhang, Zhi and Wang, Binghui and Ma, Hua and Abuadbba, Alsharif and Xue, Minhui and Fu, Anmin and Nepal, Surya},
  booktitle={2024 IEEE Symposium on Security and Privacy (SP)},
  year={2024},
  organization={IEEE}
}

Our code is implemented and tested on PyTorch. Following packages are used by our code.

• troch==1.10.1
• torchvision==0.11.2
• numpy==1.22.0
• python-Levenshtein==0.12.2

Step 1 of the offline training phase.

python Step1_Network_Structure_Recovery/train.py

Step 2 of the offline training phase.

python Step2_Layer-wise_Hyperparameter_Inferring/train.py