README.md

Snoopy: Sniffing Your Smartwatch Passwords via Deep Sequence Learning - ACM UbiComp 2018.

Introduction

This is the code and dataset used by Snoopy, an attack system for password inference on smartwatch.

Data

Download the data through this Dropbox link. Unzip the downloaded file in the project directory and check the following subfolders:

1. train: > 33,000 labelled motion samples from 147 common swiped pattern locks. Used for network training.
2. test: > 1,500 samples, containing both seen (50) and unseen (64) pattern locks during training.
3. val: > 3,800 labelled motion samples from 61 pattern locks for model selection.

Dependency

Our code has been tesed on Keras 2.0.8 with tensorflow-gpu==1.9.0 as backend. Install required dependency as per the following setps.

1. Create the py27snoopy Conda environment: conda env create -f environment.yaml.
2. Install the specific version of recurrentshop from this fork
3. Go to this fork and follow its instruction to install seq2seq.

Run the code

• FIRST: Change the config file config.ini to decide network params and regularization strategies.

• To train the attention based lstm model:

python train_att_seq.py

• To train the standard lstm model:

python train_seq.py

• To test the model:

python test.py *model_name*.hdf5

For example, python test.py model_attention_32_0.005_200_33336_0.1_2.hdf5. There are some pre-baked model examples docked in the model directory.

Citation

If you find this repository and our data useful, please cite our paper

@article{lu2018snoopy,
  title={Snoopy: Sniffing your smartwatch passwords via deep sequence learning},
  author={Lu, Chris Xiaoxuan and Du, Bowen and Wen, Hongkai and Wang, Sen and Markham, Andrew and Martinovic, Ivan and Shen, Yiran and Trigoni, Niki},
  journal={Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies},
  volume={1},
  number={4},
  pages={152},
  year={2018},
  publisher={ACM}
}

Acknowledgements

This code partially builds on Seq2Seq.