These tools are based on my thesis that is available for free on:
https://www.utupub.fi/handle/10024/177844
A library containing data anonymization tools for R.
# 1. Install "devtools" package on R
install.packages("devtools")
# 2. Install the "anon" package on R
devtools::install_github("rajalah71/anon") # or /anon@paketti
# 3. Load the "anon" library
library(anon)The package contains 4 functions to anonymize your data:
The package also contains functions to measure the success of the anonymization on three levels: privacy, utility and similarity.
data("iris")
kanon = kAnon(iris, k=5, quasiIdentifiers = c("Sepal.Length", "Sepal.Width", "Petal.Length", "Petal.Width", "Species"))[1:10, ]kanon## Sepal.Length Sepal.Width Petal.Length Petal.Width Species
## 1 4.844000 3.248000 1.3840 0.1880 setosa
## 2 4.844000 3.248000 1.3840 0.1880 setosa
## 3 4.844000 3.248000 1.3840 0.1880 setosa
## 4 4.844000 3.248000 1.3840 0.1880 setosa
## 5 4.844000 3.248000 1.3840 0.1880 setosa
## 6 5.143750 3.537500 1.5625 0.3625 setosa
## 7 5.143750 3.537500 1.5625 0.3625 setosa
## 8 5.211111 3.733333 1.5000 0.2000 setosa
## 9 4.844000 3.248000 1.3840 0.1880 setosa
## 10 4.844000 3.248000 1.3840 0.1880 setosa
-
This uses all columns as quasi identifiers by default and operates on them with mean on numeric columns and with mode on categorical ones.
-
You can define your own generalization functions and on what columns the function will operate on. These functions must be on a list with item names matching the names of the quasi-identifier columns. The columns will be worked on the order you name them, or by cardinality on descending order if left unnamed.
ldiv = lDiversity(iris, sensitiveAttributes = "Species", quasiIdentifiers = c("Sepal.Length", "Sepal.Width", "Petal.Length", "Petal.Width"), l=2)[1:10, ] ldiv## Sepal.Length Sepal.Width Petal.Length Petal.Width Species
## 1 5.036538 3.4 1.601923 0.3096154 setosa
## 2 5.036538 3.4 1.601923 0.3096154 setosa
## 3 5.036538 3.4 1.601923 0.3096154 setosa
## 4 5.036538 3.4 1.601923 0.3096154 setosa
## 5 5.036538 3.4 1.601923 0.3096154 setosa
## 6 5.036538 3.4 1.601923 0.3096154 setosa
## 7 5.036538 3.4 1.601923 0.3096154 setosa
## 8 5.036538 3.4 1.601923 0.3096154 setosa
## 9 5.036538 3.4 1.601923 0.3096154 setosa
## 10 5.036538 3.4 1.601923 0.3096154 setosa
- Works similiarly to kAnon, you just need to specify the sensitive attribute(s). If your data has multiple sensitive attributes with too high cardinality, one can use the ‘sensitive_generalizer’ function to reduce the cardinality of those variables. This, however, will negatively impact the utility of the data.
rsa = encrypt(iris)[1:10, ]rsa## Sepal.Length Sepal.Width Petal.Length Petal.Width Species
## 1 0.5062727 0.85605128 0.1306362 0.95801477 0
## 2 0.4924734 0.11105757 0.1306362 0.95801477 0
## 3 0.1035438 0.21286712 0.3367670 0.95801477 0
## 4 0.1131051 0.06792302 0.2303505 0.95801477 0
## 5 0.1469582 0.97099976 0.1306362 0.95801477 0
## 6 0.1432559 0.67730414 0.7074619 1.00000000 0
## 7 0.1131051 0.01329477 0.1306362 0.06332291 0
## 8 0.1469582 0.01329477 0.2303505 0.95801477 0
## 9 0.2352112 0.94542718 0.1306362 0.95801477 0
## 10 0.4924734 0.06792302 0.2303505 0.71684865 0
- Encrypts the data using a new RSA key = 2048 bits by default. Do not distribute the public key.
# General usage:
spectral(iris, your_anon_function, on_matrices = "U")
- Anonymizes data on the spectral basis provided by SVD,
$M = UDV'$ . Supports any anonymization function$f: R^{a \times b} \to R^{a \times b}$ . In this package, helper functions ‘cell_swap’ and ‘sensitive_noise’ are provided for permutations and additive noise respectively.
# Example usage, spectral permutation:
# Performs column permutations on each column of the matrix U
swap = spectral(iris, cell_swap, on_matrices = "U")swap[1:10, ]## Sepal.Length Sepal.Width Petal.Length Petal.Width Species
## 1 5.906487 2.375044 5.527025 2.1438910 virginica
## 2 5.847684 2.546313 4.365635 1.5204523 versicolor
## 3 5.306746 2.705059 4.028581 1.3372955 virginica
## 4 5.979122 3.195010 4.241218 1.3209798 virginica
## 5 5.026914 3.566041 1.349429 0.2506105 setosa
## 6 6.269000 2.947778 3.900450 1.5309855 virginica
## 7 7.114459 3.143902 5.825558 1.3976396 versicolor
## 8 6.035461 2.587460 5.519373 1.9293927 versicolor
## 9 6.238483 1.997127 5.770913 2.0936210 versicolor
## 10 4.918707 3.244483 1.459265 0.1719687 setosa
# Example usage spectral noise:
# Adds random noise on the columns of the matrix UD
noise = spectral(iris, sensitive_noise, on_matrices = "UD")noise[1:10, ]## Sepal.Length Sepal.Width Petal.Length Petal.Width Species
## 1 5.337824 2.8861675 5.1825368 3.4646994 virginica
## 2 7.791283 -0.2818075 4.9768479 -0.5521176 virginica
## 3 -2.445947 2.9421279 -1.6576780 1.4666112 virginica
## 4 5.748367 7.0424958 -1.8629956 4.4918546 setosa
## 5 1.315619 4.8930406 -10.1622020 -5.1917895 setosa
## 6 3.664768 3.1356003 2.2826870 1.5216045 virginica
## 7 4.229952 4.9171275 -0.4297981 -0.8479759 setosa
## 8 6.889078 -0.3392106 11.2470468 5.7660687 virginica
## 9 3.476335 -2.5037215 9.0381173 4.8878416 virginica
## 10 4.459015 2.6773951 8.0109034 3.1565152 virginica
# Re-identification rate
ri_rates = reidentification_rate(list("original" = iris, "anon" = swap), quasiIdentifiers = c("Sepal.Length", "Sepal.Width", "Petal.Length", "Petal.Width", "Species"))
lapply(ri_rates, median)## $original
## [1] 0.9933333
##
## $anon
## [1] 0.006666667
- The re-identification rate has dropped significantly after the anonymization.
prediction_plot(iris, k=5, swap)
- Also check the utility (performance on a model of interest, for example) and similarity (of means and variances, for example) of the anonymized data against those of the original data.
# Utility
trainIndices = sample(c(TRUE,FALSE), nrow(iris), replace = TRUE, prob = c(0.7, 0.3))
train = iris[trainIndices, ]
test = iris[!trainIndices, ]
anon_train = spectral(train, sensitive_noise, on_matrices = "UD")
lm = lm(Sepal.Length ~ Sepal.Width + Petal.Length + Petal.Width + Species, data = train)
lm_anon = lm(Sepal.Length ~ Sepal.Width + Petal.Length + Petal.Width + Species, data = anon_train)
testRsquared_multimodel(list("Original" = lm, "Anon" = lm_anon), test, response = "Sepal.Length")## names results
## [1,] "Original" "0.866588998862304"
## [2,] "Anon" "0.799721276792281"
- The coefficient of determination on test data has dropped after the anonymization, meaning that the utility of the data has dropped.
# Similarity
mediansAll(list("Original" = iris, "Anon" = swap))## [1] "Means:"
## $Medians
## Anon
## -8.658583e-17
##
## $MADs
## Anon
## 1.843057e-16
##
## [1] "Vars:"
## $Medians
## Anon
## 0.01042568
##
## $MADs
## Anon
## 0.009326122
##
## [1] "Cors:"
## $Medians
## Anon
## 0.02565579
##
## $MADs
## Anon
## 0.01554227
- The means, variances and correlations have changed after the anonymization.