AddShare: A Privacy-Preserving Approach for Federated Learning

Centralized machine learning methods that depend on data from multiple sources have faced serious privacy issues. Federated learning (FL), which enables decentralized machine learning by training models on local devices while keeping data private and simply sharing model weights with a central server, has therefore come to be recognized as a promising alternative. However, maintaining the privacy of training data does not protect privacy as training data can be inferred from model weights. To overcome this difficulty, we propose AddShare an FL system that protects the privacy of the local model weights while allowing the computation of a global model. Leveraging state-of-the-art techniques, AddShare uses additive secret-sharing, providing a simple yet efficient method to safeguard sensitive information without compromising predictive accuracy. Moreover, additional components are integrated to ensure lower computational costs and increase privacy. We conducted extensive experiments across multiple datasets that yielded very promising results for AddShare. AddShare did not adversely impact model accuracy compared to the widely-used FedAvg algorithm. Simultaneously, the privacy of the models is ensured and the computational cost is reduced through the implementation of groups while using a single aggregating server, a capability not available in other related solutions.