vankesteren/privreg

Privacy-preserving regression for vertically partitioned data

Private regression using block coordinate descent

installation

install.packages('privreg', repos = c('https://vankesteren.r-universe.dev', 'https://cloud.r-project.org'))

usage

library(privreg)

# create test data
set.seed(45)
S <- cov2cor(rWishart(1, 10, diag(10))[,,1])
X <- cbind(MASS::mvrnorm(1000, rep(0, 10), S), rbinom(100, 1, 0.1))
b <- runif(11, -1, 1)
y <- X %*% b + rnorm(100, sd = sd(X %*% b))

# vertically partition test data
alice_data <- data.frame(y, X[, 1:5])
bob_data   <- data.frame(y, X[, 6:11])

# create alice and bob locations
alice <- PrivReg$new(
  formula   = y ~ .,
  data      = alice_data,
  family    = gaussian(),
  intercept = TRUE,
  verbose   = TRUE,
  name      = "alice",
  crypt_key = "pre-shared-key-123"
)

bob <- PrivReg$new(
  formula   = y ~ .,
  data      = bob_data,
  family    = gaussian(),
  intercept = FALSE,
  verbose   = TRUE,
  name      = "bob  ",
  crypt_key = "pre-shared-key-123"
)

# connect the session
alice$listen()
bob$connect(url = "127.0.0.1")

# ...

# estimate the model
alice$estimate()

# ...

# disconnect
alice$disconnect()

# compare results to glm()
summary(glm(y ~ X))
alice$summary()
bob$summary()

> summary(glm(y ~ X))

Call:
glm(formula = y ~ X)

Deviance Residuals: 
     Min        1Q    Median        3Q       Max  
-2.17875  -0.73828  -0.05085   0.72709   2.45231  

Coefficients:
            Estimate Std. Error t value Pr(>|t|)    
(Intercept) -0.10280    0.03419  -3.007 0.002708 ** 
X1           0.86003    0.19318   4.452 9.48e-06 ***
X2          -0.26996    0.12100  -2.231 0.025903 *  
X3           0.85155    0.10760   7.914 6.65e-15 ***
X4          -1.07249    0.13248  -8.096 1.67e-15 ***
X5           0.36944    0.11597   3.186 0.001490 ** 
X6           1.05666    0.16451   6.423 2.07e-10 ***
X7           0.37315    0.11277   3.309 0.000971 ***
X8          -0.61272    0.19582  -3.129 0.001806 ** 
X9           0.44891    0.10730   4.184 3.12e-05 ***
X10         -0.69175    0.18520  -3.735 0.000198 ***
X11         -0.65930    0.12914  -5.105 3.96e-07 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

(Dispersion parameter for gaussian family taken to be 1.080014)

    Null deviance: 2142.7  on 999  degrees of freedom
Residual deviance: 1067.1  on 988  degrees of freedom
AIC: 2928.8

Number of Fisher Scoring iterations: 2

> alice$summary()

Privacy-preserving GLM
----------------------

family:      gaussian
formula:     y ~ .
iterations:  399

Coefficients:
             Estimate Std. Error     2.5%     97.5% t value  Pr(>|t|)    
(Intercept) -0.102797   0.034189 -0.16989 -0.035705 -3.0067  0.002708 ** 
X1           0.860031   0.193180  0.48094  1.239122  4.4520 9.478e-06 ***
X2          -0.269965   0.121005 -0.50742 -0.032509 -2.2310  0.025903 *  
X3           0.851555   0.107599  0.64041  1.062703  7.9142 6.653e-15 ***
X4          -1.072488   0.132478 -1.33246 -0.812517 -8.0956 1.666e-15 ***
X5           0.369441   0.115970  0.14186  0.597018  3.1857  0.001490 ** 
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

> bob$summary()

Privacy-preserving GLM
----------------------

family:      gaussian
formula:     y ~ .
iterations:  400

Coefficients:
   Estimate Std. Error     2.5%    97.5% t value  Pr(>|t|)    
X1  1.05666    0.16451  0.73382  1.37949  6.4229 2.074e-10 ***
X2  0.37315    0.11277  0.15185  0.59444  3.3089 0.0009705 ***
X3 -0.61272    0.19582 -0.99700 -0.22844 -3.1289 0.0018059 ** 
X4  0.44891    0.10730  0.23835  0.65947  4.1837 3.123e-05 ***
X5 -0.69175    0.18520 -1.05519 -0.32832 -3.7351 0.0001984 ***
X6 -0.65930    0.12914 -0.91272 -0.40588 -5.1053 3.961e-07 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

privreg's Issues

Running the code

Hello all! I am in the process of using your paper's ideas to write code in python. I tried to run the example you have in the README but don't actually get the same results. It looks like it's only running for 0 iterations locally. I didn't make any changes from your example. Do you have any idea about why this may be? Thank you in advance!

Make data transfer for profiling more efficient

Stop sending profile predictions for profile probes that have already converged. This will speed up the whole thing drastically when P is large.