SafeML: A Privacy-Preserving Byzantine-Robust Framework for Distributed Machine Learning Training
Meghdad Mirabi, René Klaus Nikiel, Carsten Binnig · 2023
This paper introduces SafeML, a distributed machine learning framework that can address privacy and Byzantine robustness concerns during model training. It employs secret sharing and data masking techniques to secure all computations, while also utilizing computational redundancy and robust confirmation methods to prevent Byzantine nodes from negatively affecting model updates at each iteration of model training. The theoretical analysis and preliminary experimental results demonstrate the security and correctness of SafeML for mode training.