Use cases

• Public reports.
  
• Public data sharing, e.g. R packages download logs from CRAN's RStudio mirror – cran-logs.rstudio.com – mask ip addresses.
• Reports or data sharing for external vendor.
• Development works can operate on anonymized PRODUCTION data.
  Manually or semi-manually populated data can often brings some new issue after migration to PRODUCTION data.
  Such anonymized PRODUCTION data can be quite handy for the devs.
  

Dependencies

suppressPackageStartupMessages({
  library(data.table)
  library(digest)
  library(knitr) # used only for post creation
})

Sample of survey data

Anonymize sensitive information in survey data, data storage in a single table.

# pretty print
kable(head(SURV))

Anonymize function

Function will calculate hashes only for unique inputs and return vector of masked inputs.
My version will use digest(x, algo="crc32") because it fits better into html tables, algo crc32 is not really secure.
Read ?digest::digest for supported algo, also consider to salt your input vector, e.g. x=paste0("prefix",x,"suffix").
Performance improvement possible using Rcpp / C: digest #2.

anonymize <- function(x, algo="crc32"){
  unq_hashes <- vapply(unique(x), function(object) digest(object, algo=algo), FUN.VALUE="", USE.NAMES=TRUE)
  unname(unq_hashes[x])
}

Anonymize survey data

We will keep city and sex fields unmasked.

# choose columns to mask
cols_to_mask <- c("name","address","postal_code")
# backup original data
SURV_ORG <- copy(SURV)
# anonymize
SURV[,cols_to_mask := lapply(.SD, anonymize),.SDcols=cols_to_mask,with=FALSE]
# pretty print
kable(head(SURV))

Why not just random data or integer sequence

When using the digest function to hide sensitive data you:

• keep rows distribution:
  aggregates by masked columns will still match to aggregates on original columns, see simple grouping below:

SURV_ORG[,.(.N,mean_age=mean(age),mean_score=mean(score)),by=.(city,postal_code)
         ][,kable(.SD)]

SURV[,.(.N,mean_age=mean(age),mean_score=mean(score)),by=.(city,postal_code)
     ][,kable(.SD)]

• keep relationships on equi joins:
  if t1.col1 == t2.col4 TRUE then also digest(t1.col1) == digest(t2.col4) TRUE
  Example in next section below.

Sample of sales data

Anonymize relational data in sales data, data normalized into SALES and CUSTOMER tables.

kable(head(SALES,4))

kable(head(CUSTOMER,2))

# join
kable(head(
  CUSTOMER[SALES]
))

# join and aggregate
kable(head(
  CUSTOMER[SALES][,.(quantity = sum(quantity),value = sum(value)),by=.(city,postal_code)]
))

Anonymize sales data

SALES[, customer_uid := anonymize(customer_uid)]
cols_to_mask <- c("customer_uid","name","address","postal_code")
CUSTOMER[,cols_to_mask := lapply(.SD, anonymize),.SDcols=cols_to_mask,with=FALSE]
setkey(CUSTOMER,customer_uid)

# preview result
kable(head(CUSTOMER,2))

kable(head(SALES,2))

# datasets will still join correctly even on masked columns
kable(head(
  CUSTOMER[SALES]
))

# also the aggregates on masked columns will match to the origin
kable(head(
    CUSTOMER[SALES][,.(quantity = sum(quantity),value = sum(value)),by=.(city,postal_code)]
))

Reproduce from Rmd

Script used to produce this post is available in the github repo (link in the page footer) as Rmd file and can be easily reproduced locally in R (required knitr or rmarkdown) to any format (md, html, pdf, docx).

# html output
rmarkdown::render("2014-11-07-Data-Anonymization-in-R.Rmd", html_document())
# markdown file used as current post
knitr::knit("2014-11-07-Data-Anonymization-in-R.Rmd")

Minimal script

Minimal script example on survey data as SURV_ORG data.table:

anonymize <- function(x, algo="crc32"){
  unq_hashes <- vapply(unique(x), function(object) digest(object, algo=algo), FUN.VALUE="", USE.NAMES=TRUE)
  unname(unq_hashes[x])
}
cols_to_mask <- c("name","address","postal_code")
SURV_ORG[, cols_to_mask := lapply(.SD, anonymize), .SDcols=cols_to_mask, with=FALSE][]