Theory for the Privacy and Security of Practical Federated Learning

Federated learning (FL) is a distributed machine learning paradigm that enables collaborative training while maintaining data privacy.

FL faces significant challenges, such as privacy breaches in the form of inference attacks and security breaches in the form of poisoning attacks.

Various mechanisms, including differential privacy and secure aggregation, have been proposed to address these challenges. However, these mechanisms come at a cost of reduced model accuracy and security.

Practical aspects of machine learning, such as regularization, quantization, model sparsification, and heterogeneity across clients, also affect the privacy and security of FL.

Despite the critical implications of FL´s privacy and security aspects in healthcare, finance, or autonomous driving, there are no works that formally quantify the privacy and security guarantees of FL.

This project aims to provide a fundamental understanding of the privacy and security of FL under practical considerations.

It will analyze the impact of the practical aspects mentioned above on the privacy and security of FL by exploiting the notion of noiseless differential privacy and tools from information theory, thus formally quantifying the privacy guarantees of FL and providing a precise characterization of its security.

The project´s groundbreaking contributions in the crucial and rapidly-evolving area of FL will pave the way for its adoption in real-world applications with privacy and security guarantees.