Search

Analyzing and processing data that are siloed and dispersed among multiple distrustful stakeholders is difficult and can even become impossible when the data are sensitive or confidential. Current data-protection and privacy regulations highly restrict the sharing and outsourcing of personal information among stakeholders that are in different jurisdictions. Sharing data is, however, required in many domains. The medical sector is a paradigmatic example: Privacy is paramount and data sharing is needed in numerous applications where data is scarce and scattered among multiple stakeholders around the world. Existing privacy-preserving solutions for federated analytics (FA) rely either (1) on data centralization or outsourcing to a limited number of entities, which incur multiple security and trust issues, or (2) on the exchange of cleartext aggregated and optionally obfuscated data, which can leak personal information or introduce bias in the final result. In this thesis, our goal is (1) to propose privacy-preserving federated solutions for exploration, and for statistical and machine-learning analyses on data held by multiple distrustful stakeholders, and (2) to analyze and evaluate the proposed systems, thus showing that they provide an efficient, secure, scalable, and accurate alternative to existing solutions for FA by proving their utility in real-world state-of-the-art biomedical studies. We rely on multiparty homomorphic encryption (MHE). MHE combines secure multiparty computation (SMC) techniques with homomorphic encryption (HE) by pooling the advantages of both SMC and HE, i.e., interactivity and flexibility, and by minimizing their disadvantages, i.e., difficulty in scaling to a large number of parties and computation complexity. First, we design UnLynx, a system that enables privacy-preserving federated data exploration on a distributed dataset held by multiple data-providers (DPs), where N-1 out of N of the nodes performing the computations can be malicious. We build interactive protocols by relying on ElGamal additive homomorphic encryption (AHE) and ensure that each untrusted-node operation can be publicly verified by means of zero-knowledge proofs (ZKPs). We then explore how statistics, e.g., standard deviation and variance, can be computed by relying on AHE and ZKPs through the design of another system named Drynx. In Drynx, we also explore how to limit the influence of an entity that inputs wrong data in the system, and we propose an efficient federated solution for correctness verification. We propose Spindle a solution for secure cooperative gradient descent on federated data that we instantiate for the privacy-preserving training and oblivious evaluation of generalized linear models. Spindle covers the entire machine-learning workflow, as it enables oblivious predictions to be performed on a trained model that remains secret. It ensures both data and model confidentiality in a passive adversarial model in which N-1 out of N DPs can collude. Finally, we demonstrate that the solutions proposed in this thesis can be efficient enablers for large-scale, sensitive, multi-site biomedical studies. We design and test, by replicating recent medical studies, secure workflows for the federated execution of computations that span from analyses with low computational complexity, such as survival analyses, to analyses with high computational complexity such as genome-wide association studies on millions of variants.

Medical studies are usually time consuming, cumbersome and extremely costly to perform, and for exploratory research, their results are also difficult to predict a priori. This is particularly the case for rare diseases, for which finding enough patients is difficult and usually requires an international-scale research. In this case, the process can be even more difficult due to the heterogeneity of data-protection regulations, making the data sharing process particularly hard., In this short paper, we propose MedCo(2) (pronounced MedCo square), a distributed system that streamlines the process of a medical study by bridging and enabling both data discovery and data analysis among multiple databases, while protecting data confidentiality and patients' privacy. MedCo(2) relies on interactive protocols, homomorphic encryption and differential privacy. It enables the privacy-preserving computations of multiple statistics such as cosine similarity and variance, and the training of machine learning models, on patients that are obliviously selected according to specific criteria among multiple databases.

A computer-implemented method and a distributed computer system (100) for privacy- preserving distributed training of a global model on distributed datasets (DS1 to DSn). The system has a plurality of data providers (DP1 to DPn) being communicatively coupled. Each data provider has a respective local model (LM1 to LMn) and a respective local training dataset (DS1 to DSn) for training the local model using an iterative training algorithm (IA). Further it has a portion of a cryptographic distributed secret key (SK1 to SKn) and a corresponding collective cryptographic public key (CPK) of a multiparty fully homomorphic encryption scheme, with the local and global model being encrypted with the collective public key. Each data provider (DP1) trains its local model (LM1) using the respective local training dataset (DS1) by executing gradient descent updates of its local model (LM1), and combining (1340) the updated local model (LM1') with the current global model (GM) into a current local model (LM1c). At least one data provider homomorphically combines at least a subset of the current local models of at least a subset of the data providers into a combined model (CM1), and updates the current global model (GM) based on the combined model. The updated global model is provided to at least a subset of the other data providers.

A computer-implemented method and a distributed computer system (100) for privacy- preserving distributed training of a global neural network model on distributed datasets (DS1 to DSn). The system has a plurality of data providers (DP1 to DPn) being communicatively coupled. Each data provider has a respective local training dataset (DS1 to DSn) and a vector of output labels (OL1 to OLn) for training the global model. Further, it has a portion of a cryptographic distributed secret key (SK1 to SKn) and a corresponding collective cryptographic public key (CPK) of a multiparty fully homomorphic encryption scheme, with the weights of the global model being encrypted with the collective public key. Each data provider (DP1) computes and aggregates, for each layer of the global model, encrypted local gradients (LG1) using the respective local training dataset (DS1) and output labels (OL1), with forward pass and backpropagation using stochastic gradient descent. At least one data provider homomorphically combines at least a subset of the current local gradients of at least a subset of the data providers into combined local gradients, and updates the weights of the current global model (GM) based on the combined local gradients.