A Cost-effective Framework for Privacy Preserving Federated Learning

Recently, Federated Learning (FL) has received significant attention in collaborative and privacy preserving model training across different application areas. FL is a part of machine learning that enables multiple data owners to collaboratively train a single model by sharing model parameters instead of their original data. However, it may introduce risks of information leakage while exchanging model parameters. To address this concern, existing FL approaches use Secure Multiparty Computation (SMC), Differential Privacy (DP), Homomorphic Encryption (HE), and various hybrid secure techniques. This increases the computational and communicational costs as well as the training time. In this paper, we propose a cost-effective and efficient federated learning model with vertical data partitioning. In our approach, we exploit the trivial distribution of dataset into private and non-private features. We use only private dataset for security enforcement whereas the complete dataset is used for training to enhance the accuracy of the model. Our approach also reduces the computational and communicational costs because it uses non-private data in the initial round and exchanges partially trained model parameters. Subsequently, it follows traditional FL approach with private data. Our framework doesn’t incorporate SMC during the initial round, whereas, it is only used during training with private data. We have evaluated the security and performance of our proposed framework and compared the same with state-of-art research.