GDI D8.6 - Report on privacy-enhancing solutions

GDI Pillar III aims to explore use cases and innovative applications for analysing genomic and clinical data, ideally supported by the infrastructure being deployed at the national nodes within Pillar II. As described in the Report on federated learning technologies (Deliverable D8.5), artificial intelligence techniques and – more specifically - federated learning is a recent trend that can be observed in this area.

Federated learning is a distributed machine learning technique in which multiple participants, which provide remote devices or siloed data centres, collaboratively train a shared machine learning model while keeping their data locally, better supporting data privacy. It enables collaborative learning from distributed data sources without sharing the original data, thus reducing privacy concerns and leveraging the aggregate knowledge available to the multiple participants.

Privacy-Enhancing Technologies (PETs) enhance the privacy of data that is being collected or processed, e.g. by using access control and consent management systems or using techniques for data anonymization or pseudonymization.

Privacy-Preserving Technologies (PPTs) are very useful in a distributed setting with multiple parties because they complement and strengthen Privacy-Enhancing Technologies (PETs) by addressing some of the core challenges of multi-party collaboration. They minimize trust requirements by allowing parties to collaborate without needing to share raw data. PPTs also reduce the amount of data that needs to be shared or centralized, which lowers the risk of breaches or misuse. PPTs can also protect intermediate results of PETs.

The Report on federated learning technologies (Deliverable D8.5) focusses on the most common form of federated learning, where the model is trained locally by each participant on its own data and model updates are sent to a central server. The central server then aggregates these updates to improve the global model, which is then sent back to the participants for further iterative training rounds. This enhances privacy-preservation to some extent. But other privacy-preserving technologies exist.

In this report, three other PPTs are assessed: Differential Privacy, Homomorphic Encryption and Secure Multiparty Computation. Three domains of use cases that can benefit from privacy-preserving technologies are considered: Genome-Wide Association Studies, Rare Diseases and Federated Machine Learning.

Genome-Wide Association Studies are a wide class of applications for analysing genomic and clinical data that should be supported by the infrastructure being deployed at the national nodes in a federated manner. The Use case demonstrator package (Deliverable D7.1) describes the more specific scenario (infectious diseases use case 1) to identify variants determining the severity of COVID-19 disease progression. This specific scenario uses the GDI Infectious Diseases data, which the 1+MG WG11 is currently generating.

The Rare Diseases use case also covers a wide range of genomic applications. The Use case demonstrator package (Deliverable D7.1) describes the more specific scenario to answer the question regarding the side effects of medications caused by some gene variants, using the B1MG Rare Diseases dataset.

The Federated Machine Learning use case is where the Report on federated learning technologies (Deliverable D8.5) focuses on. The described privacy-preserving technologies can be used in combination with federated machine learning, enhancing privacy. Federated machine learning can be used in a very wide range of machine learning genomic applications. The Use case demonstrator package (Deliverable D7.1) and Report on federated learning technologies (Deliverable D8.5) describe some data-driven models for Cancer Research that can be built using federated machine learning.