Differential privacy in multi-task learning approaches

This package provides an implementation of differential privacy in several multi-task learning (MTL) approaches for classification and regression problems, respectively. Basically, it offers the non-federated versions of the approaches implemented in the R packages dsMTLBase (https://github.com/transbioZI/dsMTLBase) and dsMTLClient (https://github.com/transbioZI/dsMTLClient) for differential privacy in federated multi-task learning methods as introduced in Schefzik et al. (2025) as an add-on to the implemention presented in Cao et al. (2022). In particular, differential privacy implementations for the three different MTL approaches MTL_L21, MTL_Trace and MTL_Net as described in Schefzik et al. (2025) are provided, for regression and classification settings, respectively.

Roman Schefzik aggregated, edited and finalized the code of the MTL and differential privacy functions, respectively. Han Cao provided the original implementation of the MTL functions. Xavier Escriba-Montagut and Juan R. Gonzalez contributed to the implemenation of the differential privacy mechanism.

install.packages("devtools")
library(devtools)
install_github("RomanSchefzik/DPMTL")

Examples for a usage of the functions can be found in the respective function documentations. Exemplarily, we here explicitly consider the usage of the MTL_L21 implementation in a classification setting.

library(DPMTL)
#creation of synthetic predictor matrices and response vectors and setting of the model hyperparameters
set.seed(24)
X0<-list(matrix(rnorm(100000),nrow=200,ncol=500),matrix(rnorm(150000),nrow=300,ncol=500))
set.seed(24)
Y0<-list(sample(c(1,-1),200,replace=TRUE),sample(c(1,-1),300,replace=TRUE))
lam0<-0.05
C0<-1
#non-private model (i.e., model without using differential privacy)
opts0<-list(init=0,w0=NULL,tol=0.01,maxIter=50,ter=2,diffPrivEpsilon=NULL,nRunsSensitAn=NULL)
model.nonprivate<-LR_MTL_L21(X=X0,Y=Y0,lam=lam0,C=C0,opts=opts0)
#example for a private model based on a privacy parameter of 0.7 and 100 runs for the involved sensitivity analyses
opts1<-list(init=0,w0=NULL,tol=0.01,maxIter=50,ter=2,diffPrivEpsilon=0.7,nRunsSensitAn=100)
set.seed(24)
model.private<-LR_MTL_L21(X=X0,Y=Y0,lam=lam0,C=C0,opts=opts1)

Cao, H., Zhang, Y., Baumbach, J., Burton, P. R., Dwyer, D., Koutsouleris, N., Matschinske, J., Marcon, Y., Rajan, S., Rieg, T., Ryser-Welch, P., Späth, J., The COMMITMENT Consortium, Herrmann, C., and Schwarz, E. (2022). dsMTL: a computational framework for privacy-preserving, distributed multi-task machine learning. Bioinformatics, 38(21), 4919-4926. DOI: 10.1093/bioinformatics/btac616

Schefzik, R., Cao, H., Rajan, S., Escriba-Montagut, X., Gonzalez, J. R., and Schwarz, E. (2025). Integrating differential privacy into federated multi-task learning algorithms in dsMTL.