Using ensemble machine learning algorithm for applications in recognizing false data injection attacks and an effective privacy-improving in smart grid

Abstract

One well-known machine learning paradigm, federated learning (FL) assists with data privacy by letting clients save raw data locally and sending only native parameters for the model to a data aggregator server to build a shared global model. Unfortunately, federated learning may be hacked by unscrupulous aggregators who use model parameters to deduce customers' training data. Most of the existing solutions to this problem rely on a non-colluded server setup, rely on a trusted third party to calculate master secret keys, or use a safe multiparty computation protocol, none of which improve efficiency when applied to repeated computations of an aggregate model. We provide a privacy-preserving cross-silo federated learning technique that is both efficient and secure. An effective privacy-preserving federated educational protocol is achieved by utilizing secret sharing only during the establishment phase and iterations if parties re-join, and by accelerating. Computation achievement via parallel computing. Our double-layer encryption Scheme does not require computing discrete logarithm. Additionally, clients are allowed to drop out and re-join throughout the training process. Theoretically and experimentally, the suggested technique achieves acceptable model utilities while providing proved anonymity vs. an honest-but-curious aggregator server. The method is implemented in smart grids for the purpose of detecting fake data injection attacks (FDIA). This outshines previous efforts by providing a safe method of cross-silo FDIA federated training that is resistant to assaults on local private data inference.