Integrating differential privacy into federated multi-task learning algorithms in dsMTL

Multi-task learning (MTL) enables simultaneous learning of related regression or classification tasks by exploiting shared information. The R package dsMTL provides a computational framework for federated MTL approaches, supporting the analysis of sensitive, individual-level data from geographically distributed data sources using the DataSHIELD platform. While the current architecture provides comprehensive data security mechanisms, these are not specifically tailored to MTL models. In particular, these models may still be vulnerable to membership inference attacks, attempting to determine whether a specific individual was included in a given training set using the model.

To further enhance the privacy-preserving capabilities of dsMTL and protect against such attacks, differential privacy using the Laplace mechanism is integrated into dsMTL as a novel optional feature. This approach aims to obscure individual-level characteristics from the model while retaining group-level differences. The differential privacy implementation is validated in both simulation studies and a case study identifying schizophrenia patients from gene expression data. For practical utility, it is crucial to find an adequate balance between the degree of privacy protection and the conservation of model performance by choosing a reasonable privacy parameter within the differential privacy mechanism.

dsMTL is open-source and available at https://github.com/transbioZI/dsMTLBase (server-side) and https://github.com/transbioZI/dsMTLClient (client-side).