Trusted Federated Learning: Towards a Partial Zero-Knowledge Proof Approach
Yannis Formery, Léo Mendiboure, Jonathan Villain, Virginie Deniau, Christophe Gransart, Stéphane Delbruel · 2025
Federated Learning (FL) offers an attractive framework for collaboratively training AI models while preserving data privacy. However, it also introduces challenges in verifying the integrity and authenticity of model updates across diverse clients. Zero-Knowledge Proofs (ZKP) provide a promising means to address these issues by verifying computations without revealing underlying data. Yet, global verification using ZKP remains computationally expensive and does not scale well. To overcome these limitations, we propose a novel approach grounded in two key principles: (a) partial verification, targeting carefully selected subsets of data, can effectively mitigate adversarial attacks; and (b) robust data verification is essential, ensuring not only the consistency of model parameters but also the authenticity of the underlying data. We highlight the potential operation of this partial verification system, discuss novel research directions, and outline strategies for a wider integration into FL architectures.