joogswastaken/bf-pprl-attack

Implementation of a frequency-based re-identification attack on Bloom filters in PPRL protocols

Frequency-based attack on Bloom filters in PPRL

This repository contains an implementation of a frequency-based re-identification attack on Bloom filters in privacy-preserving record linkage protocols. The attack was first described by Christen et al.¹ and discussed on my personal website in a series dedicated to Bloom-filter-based PPRL.

How to use

You will need a frequency table of values you want to mask using Bloom filters. You can find an example in the data directory using the top 1k first names in Germany². The first column must contain values and the second column their respective absolute frequencies. The table must be CSV-formatted.

Using this project assumes you have Poetry installed. Run poetry install in the root of this repository, then drop into a virtual environment using poetry shell.

To perform the attack the same way the authors did, you will need to compute the amount of hash values k. Choose a filter size m (e.g. 256) and token size q (e.g. 2) and run the following script.

$ python compute_optimal_k.py data/german-names.csv -m 256 -q 2
24.19163983958364

In this example, k should be 24. Next, generate a list of CLKs based on the frequency information of your word list. It's advisable that you create an output directory first, e.g. using mkdir -p out. Select an amount of CLKs to generate, e.g. 1m, then run the following script with your previously selected value k.

$ python generate_bf.py data/german-names.csv out/german-names-masked.csv -n 1000000 -q 2 -m 256 -k 24

Finally, run the attack with the following script. You can enable CSV output with the --stdout-csv flag which will print the amount of exact matches, potential matches, false matches and no matches as comma-separated values. The output file contains the detailed guesses for each CLK.

$ python perform_attack.py data/german-names.csv out/german-names-masked.csv out/german-names-guess.csv -q 2 
TOTAL WORD COUNT:  1000
Exact matches:     3
Potential matches: 0
False matches:     81
No matches:        916

References

Footnotes

1. Christen, Peter, et al. "Efficient cryptanalysis of bloom filters for privacy-preserving record linkage." Pacific-Asia Conference on Knowledge Discovery and Data Mining. Springer, Cham, 2017. ↩

2. Taken from Forebears' "Most Common Last Names In Germany" (URL, Archive) ↩