AEPPFL: Accurate and Efficient Privacy Protection Federal Learning in Industrial IoT

Abstract

Machine learning in the application of Industrial Internet of Things (IIoT) has great potential. However, traditional machine learning methods require collecting user data for centralized training, which may disclose sensitive information of enterprises. To solve this problem, federated learning (FL) technology with distributed training has emerged. However, FL is not infallible, as attackers retain the capability to launch inference attacks on the local models, thereby deducing sensitive information about the original data. To address this issue, this article proposes an accurate and efficient FL solution for IIoT (AEPPFL). It employs masking techniques to safeguard local model parameters and encrypts the masks using lightweight keys generated from the computational Diffie-Hellman (CDH) problem. Without obtaining any precise local models from clients, the server can still perform accurate aggregation. Furthermore, this solution operates independently of secure channels and eliminates the need for shared keys. It can protect clients' local model parameters from inference attacks with minimal precision loss while maintaining low computation and communication overheads. For example, our proposed method demonstrates significant improvements over recent state-of-the-art approaches, achieving at least a 42.3x reduction in runtime when the number of clients is 500 and the data vector size is 50K.