The problem FederatedFHE solves

FederatedFhe leverages the power of Fully Homomorphic Encryption (FHE) to address security and privacy concerns in Federated Learning (FL). By performing aggregation on encrypted models, FLuFHE prevents the aggregator from accessing the actual model parameters, protecting against both inference attacks and stealth attacks.

Federated Learning
Federated Learning is a decentralized machine learning approach that enables training models across multiple edge devices or servers without sharing raw data. It preserves privacy by keeping data local, allowing collaborative model improvement while maintaining data confidentiality.

Issues with Federated Learning
Model Poisoning: Malicious participants may inject biased or misleading data during the federated learning process to compromise the integrity of the collaborative model.

Privacy Leakage: Attackers might exploit model updates to glean information about individual participants, compromising the privacy of sensitive data.

FederatedFhe - improving Federated Learning using fully homomorphic encryption
Utilizing Fully Homomorphic Encryption (FHE) in Federated Learning models brings significant advantages. It ensures privacy preservation by allowing computations on encrypted data, enabling secure collaboration without exposing raw information. FHE compliance supports regulatory requirements, while also increasing data utility and facilitating distributed learning in untrusted environments. The reduced communication overhead further enhances the efficiency of Federated Learning processes. Overall, FHE strengthens the privacy, security, and collaborative aspects of Federated Learning models.

Challenges I ran into

Makinging FHE possible models was hard but i found Concrete ML by Zama and was able to make simple model as well as harder ones like mnist model FHE enabled and then moved there aggregation to fhEVM on Fhenix