Notes

• Generally, every attack is run with python .
• The loaders.py file is a quasi-library of useful functions and is necessary for every attack.
• Most attacks run in Python3.

Data set

Each attack takes in elements from the data set in DATA_LOC. X-Y elements are considered to be instances of site X and are monitored. Integer elements Z without a hyphen are considered to be non-monitored elements in the open world; each should come from a different site. X, Y should start from 0.

The input files should be well-filtered, i.e. it should not have missing or empty elements. There should be at least 75 cells in each instance.

Data is in the following format: Each line is a pair time\tpacketsize. Packetsize is positive if outgoing and negative if incoming In the cell format, |packetsize| = 1.

Options

Each attack needs options to set its input/output locations and other parameters.

Here is a sample Options file. This works with this small data set (unzip it directly).

Common options are:

• CLOSED_SITENUM: Number of websites in the closed world.
• CLOSED_INSTNUM: Number of instances for each website in the closed world.
• OPEN_INSTNUM: Number of instances in the open world.
• DATA_LOC: Location of the data set.
• OUTPUT_LOC: Where to output files to.
• DATA_TYPE: Extension for the data files.
• FOLD_NUM: For ten-fold testing, this is the number (0-9) of the fold that is being tested.
• FOLD_MODE: For ten-fold testing, this is which type of fold to use.
• CORE_NAME: Most attacks are embarrasingly parallelizable. This sets the name of the core to avoid collision.

Attacks

kFP (Hayes and Danezis USENIX 2016)

Download all files:

• Loader
• Python wrapper

CUMUL (Panchenko et al. NDSS 2016)

First, compile libsvm to obtain svm-train and svm-predict. The code assumes svm-train and svm-predict are in the same folder.

Download all files:

• Loader
• Python wrapper

The performance depends heavily on the choice of SVM_C_LOG and SVM_G_LOG. It may freeze if those are not chosen correctly.

kNN (Wang et al. USENIX 2014)

Download all files:

• Loader
• Python wrapper
• Feature learner
• Classifier - g++ flearner.cpp -o flearner

Run with python3 Wa-kNN.py options-temp. The accuracy output of flearner.cpp shows tp, fp, p, and f. More accurate output is shown in the .score file in OUTPUT_LOC. Performance is heavily based on NEIGHBOURS_NUM, and somewhat based on ROUND_NUM and RECOPOINTS_NUM all in flearner.cpp; edit those directly to test. Hyperparameter selection is not covered in the original work.

OSAD (Cai et al. CCS 2012, etc.)

First, compile libsvm to obtain svm-train and svm-predict. The code assumes svm-train and svm-predict are in the same folder.

Download all files:

• Loader
• Ca-OSAD Python wrapper
• Classifier - mpiCC clLev.cpp -o clLev, then mpirun -n CORE_TOTAL ./clLev options-XX
• Stratifier - g++ clgen_stratify.cpp -o clgen_stratify

Run clLev first, then clgen_stratify, then Ca-OSAD. Warning: clLev is very computationally heavy and you should be prepared to use 100+ cores on it even for a relatively small data set; this is why it's written for mpi by default.

LEV_METHOD = 1 is required in the options file for Ca-OSAD. The code can also run the Lu version, or the two Wang versions by changing LEV_METHOD.

Feature-based SVM (Panchenko et al. WPES 2011)

First, compile libsvm to obtain svm-train and svm-predict. The code assumes svm-train and svm-predict are in the same folder.

Download all files:

• Loader
• Python wrapper

The performance depends heavily on the choice of SVM_C_LOG and SVM_G_LOG. It may freeze if those are not chosen correctly.

Old attacks