nunesgh / bvm-library Goto Github PK

In the particular case where the secret is a function of the QIDs, the deterministic leakage is being computed as zero. This bug is due to a wrong condition that should be triggered only when prior is 1:

variables['CA'][attribute].update(d = variables['CA'][attribute]['d']/dataset_size)
if variables['CA'][attribute]['d'] == 1:
    variables['CA'][attribute].update(d = variables['CA'][attribute]['d'] - 1)

Before the first line, variables['CA'][attribute]['d'] corresponds to the sum of the cardinalities of the partitions induced by the QIDs in which all secret values are the same. This count is going to be equal to the size of the dataset when prior is 1 (i.e., when there is only one possible value for the secret in the whole dataset), but also, more generally in the case of a secret that is a function of the QIDs, since then all partitions have a unique secret value and the sum of their cardinalities is the size of the dataset.

The library does not support NA or NaN values and those must be manually filled using the pandas .fillna() method before calling the BVM() class, e.g. when creating the pandas DataFrame.

Problem: The histogram provided by the package (for all attacks) is rounding probabilities with <0.5% to 0%, but for any attack the probability can never be 0. The probabilities should not be rounded.

Improvement: Each probability is being rounding before generating the histogram. The histogram could be given without any rounding in the following way:

1. Set a histogram with 100 bins: (0,1], (1,2], ..., (99,100]
2. Each probability could be placed in the corresponding bin. For instance, 0.4375 would be placed in the 43th bin, (43,44].

As the histogram is already a dictionary, it could be given with the interval of bins as the labels. For example:

{
"(0,1]": probability_1,
"(1,2]": probability_2,
...
"(99,100]": probability_100,
}