This repository contains the accompanying code for the paper "PAcT: Detecting and Classifying Privacy Behaviors of Android Applications".
This repository requires conda. Install the required packages using the script in install_software.sh. This will create a conda environment called exp and then install the necessary requirements.
The dataset folder contains all the dataset files. Each dataset for RQ1 (see paper for details) is named in the format of [practice|purpose]_<dataset-name>.txt. For example, practice_collecting.txt and purpose_functionality.txt. Similarly, dataset files for RQ2 are named in the format of [practice/purpose]_<N>_hop_<dataset-name>.txt. For example practice_1_hop_collecting.txt and purpose_1_hop_functionality.txt. The models folder in dataset contains the embedding model.
The code folder contains the code to train the models. Each python file trains the model with each individual dataset and are named similar to the dataset files in the format of final_exp_[practice|purpose]_<dataset-name>.py. For example, final_exp_practice_1_hop_functionality.py. All the trained weights for each experiment are saved in the code/experiment_results/<experiment-name> folder. For example code/experiment_results/Practice_Collection/_model.pt.
The .config folder contains the configuration for each python file and can be tweaked.
To run all experiments for RQ1 and RQ2, use the scripts run_experiment_rq_1.sh and run_experiment_rq_2.sh. To save confusion matrices, change EVAL=0 to EVAL=1 in the scripts. Each experiment run will create a new folder in the experiment_results according to the name of the experiment and save the logs and tensorboard event files there. Each run will overwrite the results from previous experiments.
Note - To reproduce the results from the paper, change EXPERIMENT=1 to EXPERIMENT=0 and EVAL=0 to EVAL=1. This will load the saved models in experiment_results directory for each experiment and save the confusion matrix from the results. Running each experiment
To change hyperparameters, such as number of epochs or learning rate, you can do so in .toml files in the .config folder.
If you have any trouble running the experiment, please email the question to - vijayanta.jain@maine.edu
If you use this code academically, please cite our work
@inproceedings{jain2022pact,
title={PAcT: Detecting and Classifying Privacy Behavior of Android Applications},
author={Jain, Vijayanta and Gupta, Sanonda Datta and Ghanavati, Sepideh and Peddinti, Sai Teja and McMillan, Collin},
booktitle={Proceedings of the 15th ACM Conference on Security and Privacy in Wireless and Mobile Networks},
pages={104--118},
year={2022}
}