/Var-CNN--DynaFlow

Code and data set for the paper "Var-CNN and DynaFlow: Improved Attacks and Defenses for Website Fingerprinting".

Primary LanguagePythonMIT LicenseMIT

Var-CNN and DynaFlow

This repository contains the code and data set for the attack model and defense described in the following paper

Var-CNN and DynaFlow: Improved Attacks and Defenses for Website Fingerprinting

Sanjit Bhat, David Lu, Albert Kwon, and Srinivas Devadas.

Citation

If you find Var-CNN or DynaFlow useful in your research, please consider citing:

@article{bhat18,
  title={{Var-CNN and DynaFlow: Improved Attacks and Defenses for Website Fingerprinting}},
  author={Bhat, Sanjit and Lu, David and Kwon, Albert and Devadas, Srinivas},
  eprint={1802.10215},
  journal={arXiv},
  year={2018}
}

Introduction

Most prior work on website fingerprinting attacks use manually extracted features, and thus are fragile to protocol changes or simple defenses. In contrast, Var-CNN uses model variations on convolutional neural networks with both packet sequence and packet timing data, In open-world settings, Var-CNN yields new state-of-the-art results of 90.9% TPR and 0.3% FPR on the Wang et al. data set.

DynaFlow is a new countermeasure based on dynamically adjusting flows to protect against website fingerprinting attacks. DynaFlow provides a similar level of security as current state-of-the-art and defeats all attacks, including our own, while being over 40% more efficient than existing defenses.

Figure 1: Var-CNN convolutional feature extractor.

Figure 2: Full Var-CNN architecture.

Dependencies

  1. Ensure that you have a functioning machine with an NVIDIA GPU inside it. Without a GPU, the model will take significantly longer to run on a CPU.
  2. To set up the relevent software stack, see the instructions here under the "Software Setup" section. For our experiments, we used Ubuntu 16.04 LTS, CUDA 8.0, CuDNN v6, Python 3.6.0, and TensorFlow 1.3.0 as a backend for Keras 2.0.8.

Var-CNN

Setup

  1. Clone this repo: git clone https://github.com/sanjit-bhat/Var-CNN--DynaFlow
  2. In the same directory, make sub-directories called preprocess and predictions. These will be used to store the randomized train/test sets and the final softmax outputs of the packet time and packet direction models.
  3. For preprocess_data.py, change data_loc to point to the location of Wang et al.'s k-NN data set. You can download Wang et al.'s data set here.
  4. For var_cnn_ensemble.py, change the 2 instances of data_dir to point to the location of the preprocess directory.
  5. For evaluate_ensemble, change prediction_dir and data_dir to point to the location of the prediction and preprocess directories.

Usage

  1. To re-create the open- and closed-world results of our paper, run python run_open_closed_world.py. This script will run open- and closed-world scenarios 10 times and output the results for Var-CNN time, direction, and ensemble at varying minimum confidence levels.
  2. By changing the parameters of the .main calls inside run_open_closed_world.py, you can re-produce our experiment with the trained unmonitored sites.

Results on Wang et al. data set

Attack Accuracy (Closed) TPR (Open) FPR (Open)
k-NN 91 ± 3 85 ± 4 0.6 ± 0.4
k-FP 91 ± 1 88 ± 1 0.5 ± 0.1
SDAE 88 86 2
Var-CNN Ensemble (conf. threshold = 0.0) 93.2 ± 0.5 93.0 ± 0.5 0.7 ± 0.1
Var-CNN Ensemble (conf. threshold = 0.5) 93.2 ± 0.5 90.9 ± 0.5 0.3 ± 0.1

DynaFlow

Setup

  1. Clone this repo: git clone https://github.com/sanjit-bhat/Var-CNN--DynaFlow
  2. Inside, make sub-directories called choices and batches.
  3. Inside batches, run 
    wget "http://people.csail.mit.edu/kwonal/batch-primes.zip"
unzip batch-primes.zip
    to download our data set and unzip it to the folder batch-primes inside batches.

Usage

  1. To re-create the defense results of our paper, run python dynaflow.py. This will run all the configurations of the defense in both the open- and closed-worlds, creating the defended traces. The condensed version of each defended trace will be saved to the choices folder. The defense results will be saved to dynaflow.results.
  2. Run python bounds_closed.py and python bounds_open.py to attain the metrics of the optimal attacker on each defended data set.
  3. To run Var-CNN on a DynaFlow-defended data set, change preprocess_data.py to point to the location of the defended data set.
  4. To run Wang et al.'s k-NN and Hayes et al.'s k-FP, download their attacks and follow their documentation.
  5. To run other configurations of your choice, change the parameters at the bottom of dynaflow.py. Make sure the paths at the bottom of bounds_closed.py and bounds_open.py correspond with those found in dynaflow.py.

Contact

sanjit.bhat (at) gmail.com

davidboxboro (at) gmail.com

kwonal (at) mit.edu

Any discussions, suggestions, and questions are welcome!